Spaces:
Running
Running
Signed-off-by: Jonathan Bnayahu <[email protected]>
results/bluebench/2025-08-03T14-35-25_evaluation_results.json
ADDED
|
@@ -0,0 +1,1282 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"environment_info": {
|
| 3 |
+
"timestamp_utc": "2025-08-03T18:35:20.055545Z",
|
| 4 |
+
"command_line_invocation": [
|
| 5 |
+
"/dccstor/jbworks/miniforge3/envs/bb/bin/unitxt-evaluate",
|
| 6 |
+
"--tasks",
|
| 7 |
+
"benchmarks.bluebench",
|
| 8 |
+
"--model",
|
| 9 |
+
"cross_provider",
|
| 10 |
+
"--model_args",
|
| 11 |
+
"model_name=azure/Azure/o1-ncf,max_tokens=1024",
|
| 12 |
+
"--output_path",
|
| 13 |
+
"./results/bluebench",
|
| 14 |
+
"--log_samples",
|
| 15 |
+
"--trust_remote_code",
|
| 16 |
+
"--batch_size",
|
| 17 |
+
"8",
|
| 18 |
+
"--verbosity",
|
| 19 |
+
"ERROR"
|
| 20 |
+
],
|
| 21 |
+
"parsed_arguments": {
|
| 22 |
+
"tasks": [
|
| 23 |
+
"benchmarks.bluebench"
|
| 24 |
+
],
|
| 25 |
+
"split": "test",
|
| 26 |
+
"num_fewshots": null,
|
| 27 |
+
"limit": null,
|
| 28 |
+
"batch_size": 8,
|
| 29 |
+
"model": "azure/Azure/o1-ncf",
|
| 30 |
+
"model_args": {
|
| 31 |
+
"max_tokens": 1024
|
| 32 |
+
},
|
| 33 |
+
"gen_kwargs": null,
|
| 34 |
+
"chat_template_kwargs": null,
|
| 35 |
+
"output_path": "./results/bluebench",
|
| 36 |
+
"output_file_prefix": "evaluation_results",
|
| 37 |
+
"log_samples": true,
|
| 38 |
+
"verbosity": "ERROR",
|
| 39 |
+
"apply_chat_template": false,
|
| 40 |
+
"trust_remote_code": true,
|
| 41 |
+
"disable_hf_cache": false,
|
| 42 |
+
"cache_dir": null
|
| 43 |
+
},
|
| 44 |
+
"unitxt_version": "1.26.5",
|
| 45 |
+
"unitxt_commit_hash": "N/A",
|
| 46 |
+
"python_version": "3.10.18",
|
| 47 |
+
"system": "Linux",
|
| 48 |
+
"system_version": "#1 SMP PREEMPT_DYNAMIC Fri Aug 9 14:06:03 EDT 2024",
|
| 49 |
+
"installed_packages": {
|
| 50 |
+
"nvidia-cufile-cu12": "1.11.1.6",
|
| 51 |
+
"triton": "3.3.1",
|
| 52 |
+
"nltk": "3.9.1",
|
| 53 |
+
"anyio": "4.9.0",
|
| 54 |
+
"tiktoken": "0.9.0",
|
| 55 |
+
"charset-normalizer": "3.4.2",
|
| 56 |
+
"nvidia-cuda-runtime-cu12": "12.6.77",
|
| 57 |
+
"pyarrow": "21.0.0",
|
| 58 |
+
"sympy": "1.14.0",
|
| 59 |
+
"mecab-ko": "1.0.1",
|
| 60 |
+
"httpcore": "1.0.9",
|
| 61 |
+
"pip": "25.2",
|
| 62 |
+
"certifi": "2025.7.14",
|
| 63 |
+
"evaluate": "0.4.5",
|
| 64 |
+
"Jinja2": "3.1.6",
|
| 65 |
+
"jsonschema-specifications": "2025.4.1",
|
| 66 |
+
"pydantic_core": "2.33.2",
|
| 67 |
+
"nvidia-cusparse-cu12": "12.5.4.2",
|
| 68 |
+
"aiosignal": "1.4.0",
|
| 69 |
+
"yarl": "1.20.1",
|
| 70 |
+
"unitxt": "1.26.5",
|
| 71 |
+
"jsonschema": "4.25.0",
|
| 72 |
+
"portalocker": "3.2.0",
|
| 73 |
+
"multiprocess": "0.70.16",
|
| 74 |
+
"nvidia-nvjitlink-cu12": "12.6.85",
|
| 75 |
+
"nvidia-cublas-cu12": "12.6.4.1",
|
| 76 |
+
"pydantic": "2.11.7",
|
| 77 |
+
"async-timeout": "5.0.1",
|
| 78 |
+
"annotated-types": "0.7.0",
|
| 79 |
+
"rouge_score": "0.1.2",
|
| 80 |
+
"contourpy": "1.3.2",
|
| 81 |
+
"nvidia-cuda-cupti-cu12": "12.6.80",
|
| 82 |
+
"matplotlib": "3.10.5",
|
| 83 |
+
"six": "1.17.0",
|
| 84 |
+
"diskcache": "5.6.3",
|
| 85 |
+
"tqdm": "4.67.1",
|
| 86 |
+
"h11": "0.16.0",
|
| 87 |
+
"zipp": "3.19.2",
|
| 88 |
+
"tzdata": "2025.2",
|
| 89 |
+
"bert-score": "0.3.13",
|
| 90 |
+
"setuptools": "80.9.0",
|
| 91 |
+
"referencing": "0.36.2",
|
| 92 |
+
"sacrebleu": "2.5.1",
|
| 93 |
+
"filelock": "3.18.0",
|
| 94 |
+
"urllib3": "2.5.0",
|
| 95 |
+
"scipy": "1.15.3",
|
| 96 |
+
"nvidia-nccl-cu12": "2.26.2",
|
| 97 |
+
"kiwisolver": "1.4.8",
|
| 98 |
+
"networkx": "3.4.2",
|
| 99 |
+
"typing-inspection": "0.4.1",
|
| 100 |
+
"sniffio": "1.3.1",
|
| 101 |
+
"rpds-py": "0.26.0",
|
| 102 |
+
"nvidia-curand-cu12": "10.3.7.77",
|
| 103 |
+
"litellm": "1.74.12",
|
| 104 |
+
"pillow": "11.3.0",
|
| 105 |
+
"datasets": "3.6.0",
|
| 106 |
+
"nvidia-cusolver-cu12": "11.7.1.2",
|
| 107 |
+
"cycler": "0.12.1",
|
| 108 |
+
"tokenizers": "0.21.4",
|
| 109 |
+
"distro": "1.9.0",
|
| 110 |
+
"idna": "3.10",
|
| 111 |
+
"MarkupSafe": "3.0.2",
|
| 112 |
+
"frozenlist": "1.7.0",
|
| 113 |
+
"pyparsing": "3.2.3",
|
| 114 |
+
"regex": "2025.7.34",
|
| 115 |
+
"jiter": "0.10.0",
|
| 116 |
+
"importlib_metadata": "8.0.0",
|
| 117 |
+
"packaging": "24.2",
|
| 118 |
+
"psutil": "7.0.0",
|
| 119 |
+
"mecab-ko-dic": "1.0.0",
|
| 120 |
+
"joblib": "1.5.1",
|
| 121 |
+
"transformers": "4.54.1",
|
| 122 |
+
"fsspec": "2025.3.0",
|
| 123 |
+
"scikit-learn": "1.7.1",
|
| 124 |
+
"dill": "0.3.8",
|
| 125 |
+
"wheel": "0.45.1",
|
| 126 |
+
"nvidia-nvtx-cu12": "12.6.77",
|
| 127 |
+
"nvidia-cusparselt-cu12": "0.6.3",
|
| 128 |
+
"lxml": "6.0.0",
|
| 129 |
+
"propcache": "0.3.2",
|
| 130 |
+
"numpy": "2.2.6",
|
| 131 |
+
"mpmath": "1.3.0",
|
| 132 |
+
"conllu": "6.0.0",
|
| 133 |
+
"safetensors": "0.5.3",
|
| 134 |
+
"requests": "2.32.4",
|
| 135 |
+
"fonttools": "4.59.0",
|
| 136 |
+
"tabulate": "0.9.0",
|
| 137 |
+
"typing_extensions": "4.12.2",
|
| 138 |
+
"absl-py": "2.3.1",
|
| 139 |
+
"accelerate": "1.9.0",
|
| 140 |
+
"nvidia-cufft-cu12": "11.3.0.4",
|
| 141 |
+
"nvidia-cuda-nvrtc-cu12": "12.6.77",
|
| 142 |
+
"click": "8.2.1",
|
| 143 |
+
"attrs": "25.3.0",
|
| 144 |
+
"exceptiongroup": "1.3.0",
|
| 145 |
+
"tenacity": "9.1.2",
|
| 146 |
+
"huggingface-hub": "0.34.3",
|
| 147 |
+
"pytz": "2025.2",
|
| 148 |
+
"aiohappyeyeballs": "2.6.1",
|
| 149 |
+
"python-dateutil": "2.9.0.post0",
|
| 150 |
+
"torch": "2.7.1",
|
| 151 |
+
"python-dotenv": "1.1.1",
|
| 152 |
+
"multidict": "6.6.3",
|
| 153 |
+
"httpx": "0.28.1",
|
| 154 |
+
"aiohttp": "3.12.15",
|
| 155 |
+
"xxhash": "3.5.0",
|
| 156 |
+
"PyYAML": "6.0.2",
|
| 157 |
+
"colorama": "0.4.6",
|
| 158 |
+
"openai": "1.98.0",
|
| 159 |
+
"threadpoolctl": "3.6.0",
|
| 160 |
+
"nvidia-cudnn-cu12": "9.5.1.17",
|
| 161 |
+
"pandas": "2.3.1",
|
| 162 |
+
"hf-xet": "1.1.5",
|
| 163 |
+
"jaraco.collections": "5.1.0",
|
| 164 |
+
"tomli": "2.0.1",
|
| 165 |
+
"backports.tarfile": "1.2.0",
|
| 166 |
+
"jaraco.context": "5.3.0",
|
| 167 |
+
"typeguard": "4.3.0",
|
| 168 |
+
"autocommand": "2.2.2",
|
| 169 |
+
"jaraco.text": "3.12.1",
|
| 170 |
+
"more-itertools": "10.3.0",
|
| 171 |
+
"platformdirs": "4.2.2",
|
| 172 |
+
"inflect": "7.3.1",
|
| 173 |
+
"jaraco.functools": "4.0.1"
|
| 174 |
+
}
|
| 175 |
+
},
|
| 176 |
+
"results": {
|
| 177 |
+
"bias": {
|
| 178 |
+
"safety_bbq_age": {
|
| 179 |
+
"accuracy": 1.0,
|
| 180 |
+
"accuracy_ci_low": 1.0,
|
| 181 |
+
"accuracy_ci_high": 1.0,
|
| 182 |
+
"score_name": "accuracy",
|
| 183 |
+
"score": 1.0,
|
| 184 |
+
"score_ci_high": 1.0,
|
| 185 |
+
"score_ci_low": 1.0,
|
| 186 |
+
"num_of_instances": 9
|
| 187 |
+
},
|
| 188 |
+
"safety_bbq_disability_status": {
|
| 189 |
+
"accuracy": 1.0,
|
| 190 |
+
"accuracy_ci_low": 1.0,
|
| 191 |
+
"accuracy_ci_high": 1.0,
|
| 192 |
+
"score_name": "accuracy",
|
| 193 |
+
"score": 1.0,
|
| 194 |
+
"score_ci_high": 1.0,
|
| 195 |
+
"score_ci_low": 1.0,
|
| 196 |
+
"num_of_instances": 9
|
| 197 |
+
},
|
| 198 |
+
"safety_bbq_gender_identity": {
|
| 199 |
+
"accuracy": 1.0,
|
| 200 |
+
"accuracy_ci_low": 1.0,
|
| 201 |
+
"accuracy_ci_high": 1.0,
|
| 202 |
+
"score_name": "accuracy",
|
| 203 |
+
"score": 1.0,
|
| 204 |
+
"score_ci_high": 1.0,
|
| 205 |
+
"score_ci_low": 1.0,
|
| 206 |
+
"num_of_instances": 9
|
| 207 |
+
},
|
| 208 |
+
"safety_bbq_nationality": {
|
| 209 |
+
"accuracy": 1.0,
|
| 210 |
+
"accuracy_ci_low": 1.0,
|
| 211 |
+
"accuracy_ci_high": 1.0,
|
| 212 |
+
"score_name": "accuracy",
|
| 213 |
+
"score": 1.0,
|
| 214 |
+
"score_ci_high": 1.0,
|
| 215 |
+
"score_ci_low": 1.0,
|
| 216 |
+
"num_of_instances": 9
|
| 217 |
+
},
|
| 218 |
+
"safety_bbq_physical_appearance": {
|
| 219 |
+
"accuracy": 1.0,
|
| 220 |
+
"accuracy_ci_low": 1.0,
|
| 221 |
+
"accuracy_ci_high": 1.0,
|
| 222 |
+
"score_name": "accuracy",
|
| 223 |
+
"score": 1.0,
|
| 224 |
+
"score_ci_high": 1.0,
|
| 225 |
+
"score_ci_low": 1.0,
|
| 226 |
+
"num_of_instances": 9
|
| 227 |
+
},
|
| 228 |
+
"safety_bbq_race_ethnicity": {
|
| 229 |
+
"accuracy": 1.0,
|
| 230 |
+
"accuracy_ci_low": 1.0,
|
| 231 |
+
"accuracy_ci_high": 1.0,
|
| 232 |
+
"score_name": "accuracy",
|
| 233 |
+
"score": 1.0,
|
| 234 |
+
"score_ci_high": 1.0,
|
| 235 |
+
"score_ci_low": 1.0,
|
| 236 |
+
"num_of_instances": 9
|
| 237 |
+
},
|
| 238 |
+
"safety_bbq_race_x_gender": {
|
| 239 |
+
"accuracy": 1.0,
|
| 240 |
+
"accuracy_ci_low": 1.0,
|
| 241 |
+
"accuracy_ci_high": 1.0,
|
| 242 |
+
"score_name": "accuracy",
|
| 243 |
+
"score": 1.0,
|
| 244 |
+
"score_ci_high": 1.0,
|
| 245 |
+
"score_ci_low": 1.0,
|
| 246 |
+
"num_of_instances": 9
|
| 247 |
+
},
|
| 248 |
+
"safety_bbq_race_x_ses": {
|
| 249 |
+
"accuracy": 1.0,
|
| 250 |
+
"accuracy_ci_low": 1.0,
|
| 251 |
+
"accuracy_ci_high": 1.0,
|
| 252 |
+
"score_name": "accuracy",
|
| 253 |
+
"score": 1.0,
|
| 254 |
+
"score_ci_high": 1.0,
|
| 255 |
+
"score_ci_low": 1.0,
|
| 256 |
+
"num_of_instances": 9
|
| 257 |
+
},
|
| 258 |
+
"safety_bbq_religion": {
|
| 259 |
+
"accuracy": 1.0,
|
| 260 |
+
"accuracy_ci_low": 1.0,
|
| 261 |
+
"accuracy_ci_high": 1.0,
|
| 262 |
+
"score_name": "accuracy",
|
| 263 |
+
"score": 1.0,
|
| 264 |
+
"score_ci_high": 1.0,
|
| 265 |
+
"score_ci_low": 1.0,
|
| 266 |
+
"num_of_instances": 9
|
| 267 |
+
},
|
| 268 |
+
"safety_bbq_ses": {
|
| 269 |
+
"accuracy": 0.4444444444444444,
|
| 270 |
+
"accuracy_ci_low": 0.1111111111111111,
|
| 271 |
+
"accuracy_ci_high": 0.7777777777777778,
|
| 272 |
+
"score_name": "accuracy",
|
| 273 |
+
"score": 0.4444444444444444,
|
| 274 |
+
"score_ci_high": 0.7777777777777778,
|
| 275 |
+
"score_ci_low": 0.1111111111111111,
|
| 276 |
+
"num_of_instances": 9
|
| 277 |
+
},
|
| 278 |
+
"safety_bbq_sexual_orientation": {
|
| 279 |
+
"accuracy": 1.0,
|
| 280 |
+
"accuracy_ci_low": 1.0,
|
| 281 |
+
"accuracy_ci_high": 1.0,
|
| 282 |
+
"score_name": "accuracy",
|
| 283 |
+
"score": 1.0,
|
| 284 |
+
"score_ci_high": 1.0,
|
| 285 |
+
"score_ci_low": 1.0,
|
| 286 |
+
"num_of_instances": 9
|
| 287 |
+
},
|
| 288 |
+
"score": 0.9494949494949495,
|
| 289 |
+
"score_name": "subsets_mean",
|
| 290 |
+
"num_of_instances": 99
|
| 291 |
+
},
|
| 292 |
+
"chatbot_abilities": {
|
| 293 |
+
"arena_hard_generation_english_gpt_4_0314_reference": {
|
| 294 |
+
"num_of_instances": 100,
|
| 295 |
+
"llama_3_70b_instruct_template_arena_hard": 0.34375,
|
| 296 |
+
"score": 0.34375,
|
| 297 |
+
"score_name": "llama_3_70b_instruct_template_arena_hard"
|
| 298 |
+
},
|
| 299 |
+
"score": 0.34375,
|
| 300 |
+
"score_name": "subsets_mean",
|
| 301 |
+
"num_of_instances": 100
|
| 302 |
+
},
|
| 303 |
+
"entity_extraction": {
|
| 304 |
+
"universal_ner_en_ewt": {
|
| 305 |
+
"num_of_instances": 100,
|
| 306 |
+
"f1_Person": 0.8695652173913043,
|
| 307 |
+
"f1_Organization": 0.7868852459016394,
|
| 308 |
+
"f1_Location": 0.723404255319149,
|
| 309 |
+
"f1_macro": 0.7932849062040309,
|
| 310 |
+
"recall_macro": 0.8116804692891649,
|
| 311 |
+
"precision_macro": 0.7786561264822135,
|
| 312 |
+
"in_classes_support": 1.0,
|
| 313 |
+
"f1_micro": 0.7922077922077922,
|
| 314 |
+
"recall_micro": 0.8133333333333334,
|
| 315 |
+
"precision_micro": 0.7721518987341772,
|
| 316 |
+
"score": 0.7922077922077922,
|
| 317 |
+
"score_name": "f1_micro",
|
| 318 |
+
"score_ci_low": 0.7030169722972905,
|
| 319 |
+
"score_ci_high": 0.848370061994058,
|
| 320 |
+
"f1_micro_ci_low": 0.7030169722972905,
|
| 321 |
+
"f1_micro_ci_high": 0.848370061994058
|
| 322 |
+
},
|
| 323 |
+
"score": 0.7922077922077922,
|
| 324 |
+
"score_name": "subsets_mean",
|
| 325 |
+
"num_of_instances": 100
|
| 326 |
+
},
|
| 327 |
+
"knowledge": {
|
| 328 |
+
"mmlu_pro_biology": {
|
| 329 |
+
"accuracy": 0.8571428571428571,
|
| 330 |
+
"accuracy_ci_low": 0.42857142857142855,
|
| 331 |
+
"accuracy_ci_high": 1.0,
|
| 332 |
+
"score_name": "accuracy",
|
| 333 |
+
"score": 0.8571428571428571,
|
| 334 |
+
"score_ci_high": 1.0,
|
| 335 |
+
"score_ci_low": 0.42857142857142855,
|
| 336 |
+
"num_of_instances": 7
|
| 337 |
+
},
|
| 338 |
+
"mmlu_pro_business": {
|
| 339 |
+
"accuracy": 0.8571428571428571,
|
| 340 |
+
"accuracy_ci_low": 0.2530277506117974,
|
| 341 |
+
"accuracy_ci_high": 1.0,
|
| 342 |
+
"score_name": "accuracy",
|
| 343 |
+
"score": 0.8571428571428571,
|
| 344 |
+
"score_ci_high": 1.0,
|
| 345 |
+
"score_ci_low": 0.2530277506117974,
|
| 346 |
+
"num_of_instances": 7
|
| 347 |
+
},
|
| 348 |
+
"mmlu_pro_chemistry": {
|
| 349 |
+
"accuracy": 0.2857142857142857,
|
| 350 |
+
"accuracy_ci_low": 0.0,
|
| 351 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 352 |
+
"score_name": "accuracy",
|
| 353 |
+
"score": 0.2857142857142857,
|
| 354 |
+
"score_ci_high": 0.7142857142857143,
|
| 355 |
+
"score_ci_low": 0.0,
|
| 356 |
+
"num_of_instances": 7
|
| 357 |
+
},
|
| 358 |
+
"mmlu_pro_computer_science": {
|
| 359 |
+
"accuracy": 0.8571428571428571,
|
| 360 |
+
"accuracy_ci_low": 0.42857142857142855,
|
| 361 |
+
"accuracy_ci_high": 1.0,
|
| 362 |
+
"score_name": "accuracy",
|
| 363 |
+
"score": 0.8571428571428571,
|
| 364 |
+
"score_ci_high": 1.0,
|
| 365 |
+
"score_ci_low": 0.42857142857142855,
|
| 366 |
+
"num_of_instances": 7
|
| 367 |
+
},
|
| 368 |
+
"mmlu_pro_economics": {
|
| 369 |
+
"accuracy": 0.8571428571428571,
|
| 370 |
+
"accuracy_ci_low": 0.42857142857142855,
|
| 371 |
+
"accuracy_ci_high": 1.0,
|
| 372 |
+
"score_name": "accuracy",
|
| 373 |
+
"score": 0.8571428571428571,
|
| 374 |
+
"score_ci_high": 1.0,
|
| 375 |
+
"score_ci_low": 0.42857142857142855,
|
| 376 |
+
"num_of_instances": 7
|
| 377 |
+
},
|
| 378 |
+
"mmlu_pro_engineering": {
|
| 379 |
+
"accuracy": 0.7142857142857143,
|
| 380 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 381 |
+
"accuracy_ci_high": 1.0,
|
| 382 |
+
"score_name": "accuracy",
|
| 383 |
+
"score": 0.7142857142857143,
|
| 384 |
+
"score_ci_high": 1.0,
|
| 385 |
+
"score_ci_low": 0.2857142857142857,
|
| 386 |
+
"num_of_instances": 7
|
| 387 |
+
},
|
| 388 |
+
"mmlu_pro_health": {
|
| 389 |
+
"accuracy": 0.5714285714285714,
|
| 390 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 391 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 392 |
+
"score_name": "accuracy",
|
| 393 |
+
"score": 0.5714285714285714,
|
| 394 |
+
"score_ci_high": 0.8571428571428571,
|
| 395 |
+
"score_ci_low": 0.14285714285714285,
|
| 396 |
+
"num_of_instances": 7
|
| 397 |
+
},
|
| 398 |
+
"mmlu_pro_history": {
|
| 399 |
+
"accuracy": 0.2857142857142857,
|
| 400 |
+
"accuracy_ci_low": 0.0,
|
| 401 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 402 |
+
"score_name": "accuracy",
|
| 403 |
+
"score": 0.2857142857142857,
|
| 404 |
+
"score_ci_high": 0.7142857142857143,
|
| 405 |
+
"score_ci_low": 0.0,
|
| 406 |
+
"num_of_instances": 7
|
| 407 |
+
},
|
| 408 |
+
"mmlu_pro_law": {
|
| 409 |
+
"accuracy": 0.7142857142857143,
|
| 410 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 411 |
+
"accuracy_ci_high": 1.0,
|
| 412 |
+
"score_name": "accuracy",
|
| 413 |
+
"score": 0.7142857142857143,
|
| 414 |
+
"score_ci_high": 1.0,
|
| 415 |
+
"score_ci_low": 0.2857142857142857,
|
| 416 |
+
"num_of_instances": 7
|
| 417 |
+
},
|
| 418 |
+
"mmlu_pro_math": {
|
| 419 |
+
"accuracy": 1.0,
|
| 420 |
+
"accuracy_ci_low": 1.0,
|
| 421 |
+
"accuracy_ci_high": 1.0,
|
| 422 |
+
"score_name": "accuracy",
|
| 423 |
+
"score": 1.0,
|
| 424 |
+
"score_ci_high": 1.0,
|
| 425 |
+
"score_ci_low": 1.0,
|
| 426 |
+
"num_of_instances": 7
|
| 427 |
+
},
|
| 428 |
+
"mmlu_pro_other": {
|
| 429 |
+
"accuracy": 0.7142857142857143,
|
| 430 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 431 |
+
"accuracy_ci_high": 1.0,
|
| 432 |
+
"score_name": "accuracy",
|
| 433 |
+
"score": 0.7142857142857143,
|
| 434 |
+
"score_ci_high": 1.0,
|
| 435 |
+
"score_ci_low": 0.2857142857142857,
|
| 436 |
+
"num_of_instances": 7
|
| 437 |
+
},
|
| 438 |
+
"mmlu_pro_philosophy": {
|
| 439 |
+
"accuracy": 0.7142857142857143,
|
| 440 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 441 |
+
"accuracy_ci_high": 1.0,
|
| 442 |
+
"score_name": "accuracy",
|
| 443 |
+
"score": 0.7142857142857143,
|
| 444 |
+
"score_ci_high": 1.0,
|
| 445 |
+
"score_ci_low": 0.2857142857142857,
|
| 446 |
+
"num_of_instances": 7
|
| 447 |
+
},
|
| 448 |
+
"mmlu_pro_physics": {
|
| 449 |
+
"accuracy": 0.7142857142857143,
|
| 450 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 451 |
+
"accuracy_ci_high": 1.0,
|
| 452 |
+
"score_name": "accuracy",
|
| 453 |
+
"score": 0.7142857142857143,
|
| 454 |
+
"score_ci_high": 1.0,
|
| 455 |
+
"score_ci_low": 0.2857142857142857,
|
| 456 |
+
"num_of_instances": 7
|
| 457 |
+
},
|
| 458 |
+
"mmlu_pro_psychology": {
|
| 459 |
+
"accuracy": 0.8571428571428571,
|
| 460 |
+
"accuracy_ci_low": 0.2530277506117974,
|
| 461 |
+
"accuracy_ci_high": 1.0,
|
| 462 |
+
"score_name": "accuracy",
|
| 463 |
+
"score": 0.8571428571428571,
|
| 464 |
+
"score_ci_high": 1.0,
|
| 465 |
+
"score_ci_low": 0.2530277506117974,
|
| 466 |
+
"num_of_instances": 7
|
| 467 |
+
},
|
| 468 |
+
"score": 0.7142857142857143,
|
| 469 |
+
"score_name": "subsets_mean",
|
| 470 |
+
"num_of_instances": 98
|
| 471 |
+
},
|
| 472 |
+
"legal": {
|
| 473 |
+
"legalbench_abercrombie": {
|
| 474 |
+
"f1_macro": 0.3633333333333333,
|
| 475 |
+
"f1_suggestive": 0.25,
|
| 476 |
+
"f1_generic": 0.5,
|
| 477 |
+
"f1_fanciful": 0.4,
|
| 478 |
+
"f1_descriptive": 0.6666666666666666,
|
| 479 |
+
"f1_arbitrary": 0.0,
|
| 480 |
+
"f1_macro_ci_low": 0.18019056979170386,
|
| 481 |
+
"f1_macro_ci_high": 0.6700601186500376,
|
| 482 |
+
"score_name": "f1_micro",
|
| 483 |
+
"score": 0.38461538461538464,
|
| 484 |
+
"score_ci_high": 0.6317641031035699,
|
| 485 |
+
"score_ci_low": 0.16,
|
| 486 |
+
"num_of_instances": 20,
|
| 487 |
+
"accuracy": 0.25,
|
| 488 |
+
"accuracy_ci_low": 0.1,
|
| 489 |
+
"accuracy_ci_high": 0.5,
|
| 490 |
+
"f1_micro": 0.38461538461538464,
|
| 491 |
+
"f1_micro_ci_low": 0.16,
|
| 492 |
+
"f1_micro_ci_high": 0.6317641031035699
|
| 493 |
+
},
|
| 494 |
+
"legalbench_corporate_lobbying": {
|
| 495 |
+
"f1_macro": 0.17647058823529413,
|
| 496 |
+
"f1_no": 0.35294117647058826,
|
| 497 |
+
"f1_yes": 0.0,
|
| 498 |
+
"f1_macro_ci_low": 0.0625,
|
| 499 |
+
"f1_macro_ci_high": 0.34407383963381494,
|
| 500 |
+
"score_name": "f1_micro",
|
| 501 |
+
"score": 0.2608695652173913,
|
| 502 |
+
"score_ci_high": 0.5714285714285714,
|
| 503 |
+
"score_ci_low": 0.09523809523809523,
|
| 504 |
+
"num_of_instances": 20,
|
| 505 |
+
"accuracy": 0.15,
|
| 506 |
+
"accuracy_ci_low": 0.05,
|
| 507 |
+
"accuracy_ci_high": 0.4,
|
| 508 |
+
"f1_micro": 0.2608695652173913,
|
| 509 |
+
"f1_micro_ci_low": 0.09523809523809523,
|
| 510 |
+
"f1_micro_ci_high": 0.5714285714285714
|
| 511 |
+
},
|
| 512 |
+
"legalbench_function_of_decision_section": {
|
| 513 |
+
"f1_macro": 0.14285714285714285,
|
| 514 |
+
"f1_conclusion": 0.0,
|
| 515 |
+
"f1_issue": 0.25,
|
| 516 |
+
"f1_decree": 0.0,
|
| 517 |
+
"f1_rule": 0.0,
|
| 518 |
+
"f1_analysis": 0.0,
|
| 519 |
+
"f1_facts": 0.75,
|
| 520 |
+
"f1_procedural history": 0.0,
|
| 521 |
+
"f1_macro_ci_low": 0.027938635003925405,
|
| 522 |
+
"f1_macro_ci_high": 0.25,
|
| 523 |
+
"score_name": "f1_micro",
|
| 524 |
+
"score": 0.23529411764705882,
|
| 525 |
+
"score_ci_high": 0.451025257708528,
|
| 526 |
+
"score_ci_low": 0.058823529411764705,
|
| 527 |
+
"num_of_instances": 20,
|
| 528 |
+
"accuracy": 0.2,
|
| 529 |
+
"accuracy_ci_low": 0.05,
|
| 530 |
+
"accuracy_ci_high": 0.4,
|
| 531 |
+
"f1_micro": 0.23529411764705882,
|
| 532 |
+
"f1_micro_ci_low": 0.058823529411764705,
|
| 533 |
+
"f1_micro_ci_high": 0.451025257708528
|
| 534 |
+
},
|
| 535 |
+
"legalbench_international_citizenship_questions": {
|
| 536 |
+
"f1_macro": 0.7013888888888888,
|
| 537 |
+
"f1_yes": 0.7777777777777778,
|
| 538 |
+
"f1_no": 0.625,
|
| 539 |
+
"f1_macro_ci_low": 0.4822715139171238,
|
| 540 |
+
"f1_macro_ci_high": 0.8740955338427088,
|
| 541 |
+
"score_name": "f1_micro",
|
| 542 |
+
"score": 0.7058823529411765,
|
| 543 |
+
"score_ci_high": 0.8648648648648649,
|
| 544 |
+
"score_ci_low": 0.4827586206896552,
|
| 545 |
+
"num_of_instances": 20,
|
| 546 |
+
"accuracy": 0.6,
|
| 547 |
+
"accuracy_ci_low": 0.4,
|
| 548 |
+
"accuracy_ci_high": 0.8,
|
| 549 |
+
"f1_micro": 0.7058823529411765,
|
| 550 |
+
"f1_micro_ci_low": 0.4827586206896552,
|
| 551 |
+
"f1_micro_ci_high": 0.8648648648648649
|
| 552 |
+
},
|
| 553 |
+
"legalbench_proa": {
|
| 554 |
+
"f1_macro": 0.9705882352941176,
|
| 555 |
+
"f1_yes": 0.9411764705882353,
|
| 556 |
+
"f1_no": 1.0,
|
| 557 |
+
"f1_macro_ci_low": 0.8333333333333333,
|
| 558 |
+
"f1_macro_ci_high": 1.0,
|
| 559 |
+
"score_name": "f1_micro",
|
| 560 |
+
"score": 0.9743589743589743,
|
| 561 |
+
"score_ci_high": 1.0,
|
| 562 |
+
"score_ci_low": 0.8571428571428571,
|
| 563 |
+
"num_of_instances": 20,
|
| 564 |
+
"accuracy": 0.95,
|
| 565 |
+
"accuracy_ci_low": 0.75,
|
| 566 |
+
"accuracy_ci_high": 1.0,
|
| 567 |
+
"f1_micro": 0.9743589743589743,
|
| 568 |
+
"f1_micro_ci_low": 0.8571428571428571,
|
| 569 |
+
"f1_micro_ci_high": 1.0
|
| 570 |
+
},
|
| 571 |
+
"score": 0.5122040789559972,
|
| 572 |
+
"score_name": "subsets_mean",
|
| 573 |
+
"num_of_instances": 100
|
| 574 |
+
},
|
| 575 |
+
"news_classification": {
|
| 576 |
+
"20_newsgroups_short": {
|
| 577 |
+
"f1_macro": 0.6497871572871573,
|
| 578 |
+
"f1_cars": 0.9090909090909091,
|
| 579 |
+
"f1_windows x": 0.75,
|
| 580 |
+
"f1_computer graphics": 0.7142857142857143,
|
| 581 |
+
"f1_atheism": 0.2857142857142857,
|
| 582 |
+
"f1_religion": 0.0,
|
| 583 |
+
"f1_medicine": 0.8571428571428571,
|
| 584 |
+
"f1_christianity": 0.8571428571428571,
|
| 585 |
+
"f1_microsoft windows": 0.8,
|
| 586 |
+
"f1_middle east": 0.6666666666666666,
|
| 587 |
+
"f1_motorcycles": 0.4444444444444444,
|
| 588 |
+
"f1_pc hardware": 0.7142857142857143,
|
| 589 |
+
"f1_mac hardware": 1.0,
|
| 590 |
+
"f1_electronics": 0.4,
|
| 591 |
+
"f1_for sale": 0.75,
|
| 592 |
+
"f1_guns": 0.4444444444444444,
|
| 593 |
+
"f1_space": 0.75,
|
| 594 |
+
"f1_cryptography": 0.4,
|
| 595 |
+
"f1_baseball": 1.0,
|
| 596 |
+
"f1_hockey": 0.8888888888888888,
|
| 597 |
+
"f1_politics": 0.36363636363636365,
|
| 598 |
+
"f1_macro_ci_low": 0.5620624688463499,
|
| 599 |
+
"f1_macro_ci_high": 0.7570253093572227,
|
| 600 |
+
"score_name": "f1_micro",
|
| 601 |
+
"score": 0.6815642458100558,
|
| 602 |
+
"score_ci_high": 0.7640449438202247,
|
| 603 |
+
"score_ci_low": 0.5781960812529161,
|
| 604 |
+
"num_of_instances": 100,
|
| 605 |
+
"accuracy": 0.61,
|
| 606 |
+
"accuracy_ci_low": 0.51,
|
| 607 |
+
"accuracy_ci_high": 0.7,
|
| 608 |
+
"f1_micro": 0.6815642458100558,
|
| 609 |
+
"f1_micro_ci_low": 0.5781960812529161,
|
| 610 |
+
"f1_micro_ci_high": 0.7640449438202247
|
| 611 |
+
},
|
| 612 |
+
"score": 0.6815642458100558,
|
| 613 |
+
"score_name": "subsets_mean",
|
| 614 |
+
"num_of_instances": 100
|
| 615 |
+
},
|
| 616 |
+
"product_help": {
|
| 617 |
+
"cfpb_product_2023": {
|
| 618 |
+
"f1_macro": 0.7978772331056343,
|
| 619 |
+
"f1_debt collection": 0.9090909090909091,
|
| 620 |
+
"f1_checking or savings account": 0.631578947368421,
|
| 621 |
+
"f1_credit reporting or credit repair services or other personal consumer reports": 0.8867924528301887,
|
| 622 |
+
"f1_mortgage": 0.6666666666666666,
|
| 623 |
+
"f1_credit card or prepaid card": 0.8,
|
| 624 |
+
"f1_payday loan or title loan or personal loan": 0.8,
|
| 625 |
+
"f1_student loan": 0.8888888888888888,
|
| 626 |
+
"f1_money transfer or virtual currency or money service": 0.8,
|
| 627 |
+
"f1_macro_ci_low": 0.6793031276678277,
|
| 628 |
+
"f1_macro_ci_high": 0.8893682569671233,
|
| 629 |
+
"score_name": "f1_micro",
|
| 630 |
+
"score": 0.84375,
|
| 631 |
+
"score_ci_high": 0.9035402051084471,
|
| 632 |
+
"score_ci_low": 0.7562033534601807,
|
| 633 |
+
"num_of_instances": 100,
|
| 634 |
+
"accuracy": 0.81,
|
| 635 |
+
"accuracy_ci_low": 0.72,
|
| 636 |
+
"accuracy_ci_high": 0.88,
|
| 637 |
+
"f1_micro": 0.84375,
|
| 638 |
+
"f1_micro_ci_low": 0.7562033534601807,
|
| 639 |
+
"f1_micro_ci_high": 0.9035402051084471
|
| 640 |
+
},
|
| 641 |
+
"cfpb_product_watsonx": {
|
| 642 |
+
"f1_macro": 0.8289604115691072,
|
| 643 |
+
"f1_mortgages and loans": 0.782608695652174,
|
| 644 |
+
"f1_credit card": 0.8571428571428571,
|
| 645 |
+
"f1_debt collection": 0.7777777777777778,
|
| 646 |
+
"f1_credit reporting": 0.7272727272727273,
|
| 647 |
+
"f1_retail banking": 1.0,
|
| 648 |
+
"f1_macro_ci_low": 0.7098713502417807,
|
| 649 |
+
"f1_macro_ci_high": 0.9150602383003535,
|
| 650 |
+
"score_name": "f1_micro",
|
| 651 |
+
"score": 0.8163265306122449,
|
| 652 |
+
"score_ci_high": 0.9072164948453608,
|
| 653 |
+
"score_ci_low": 0.6839680616837706,
|
| 654 |
+
"num_of_instances": 50,
|
| 655 |
+
"accuracy": 0.8,
|
| 656 |
+
"accuracy_ci_low": 0.66,
|
| 657 |
+
"accuracy_ci_high": 0.9,
|
| 658 |
+
"f1_micro": 0.8163265306122449,
|
| 659 |
+
"f1_micro_ci_low": 0.6839680616837706,
|
| 660 |
+
"f1_micro_ci_high": 0.9072164948453608
|
| 661 |
+
},
|
| 662 |
+
"score": 0.8300382653061225,
|
| 663 |
+
"score_name": "subsets_mean",
|
| 664 |
+
"num_of_instances": 150
|
| 665 |
+
},
|
| 666 |
+
"qa_finance": {
|
| 667 |
+
"fin_qa": {
|
| 668 |
+
"num_of_instances": 100,
|
| 669 |
+
"program_accuracy": 0.22,
|
| 670 |
+
"score": 0.22,
|
| 671 |
+
"score_name": "program_accuracy",
|
| 672 |
+
"execution_accuracy": 0.22,
|
| 673 |
+
"program_accuracy_ci_low": 0.15,
|
| 674 |
+
"program_accuracy_ci_high": 0.3,
|
| 675 |
+
"score_ci_low": 0.15,
|
| 676 |
+
"score_ci_high": 0.3,
|
| 677 |
+
"execution_accuracy_ci_low": 0.14,
|
| 678 |
+
"execution_accuracy_ci_high": 0.31
|
| 679 |
+
},
|
| 680 |
+
"score": 0.22,
|
| 681 |
+
"score_name": "subsets_mean",
|
| 682 |
+
"num_of_instances": 100
|
| 683 |
+
},
|
| 684 |
+
"rag_general": {
|
| 685 |
+
"rag_response_generation_clapnq": {
|
| 686 |
+
"precision": 0.42816457156991794,
|
| 687 |
+
"recall": 0.4569091793970677,
|
| 688 |
+
"f1": 0.4124340232392918,
|
| 689 |
+
"precision_ci_low": 0.38847540220366156,
|
| 690 |
+
"precision_ci_high": 0.4624925368575603,
|
| 691 |
+
"recall_ci_low": 0.41415283802156605,
|
| 692 |
+
"recall_ci_high": 0.4977227170242366,
|
| 693 |
+
"f1_ci_low": 0.37251708583536297,
|
| 694 |
+
"f1_ci_high": 0.44397312360149926,
|
| 695 |
+
"score_name": "f1",
|
| 696 |
+
"score": 0.4124340232392918,
|
| 697 |
+
"score_ci_high": 0.44397312360149926,
|
| 698 |
+
"score_ci_low": 0.37251708583536297,
|
| 699 |
+
"num_of_instances": 100,
|
| 700 |
+
"correctness_f1_bert_score.deberta_large_mnli": 0.6425585800409317,
|
| 701 |
+
"correctness_recall_bert_score.deberta_large_mnli": 0.652276462316513,
|
| 702 |
+
"correctness_precision_bert_score.deberta_large_mnli": 0.6418503573536873,
|
| 703 |
+
"faithfullness_f1_token_overlap": 0.29098835965027586,
|
| 704 |
+
"faithfullness_recall_token_overlap": 0.21164244081926975,
|
| 705 |
+
"faithfullness_precision_token_overlap": 0.5818246023074958,
|
| 706 |
+
"correctness_f1_token_overlap": 0.4124340232392918,
|
| 707 |
+
"correctness_recall_token_overlap": 0.4569091793970677,
|
| 708 |
+
"correctness_precision_token_overlap": 0.42816457156991794
|
| 709 |
+
},
|
| 710 |
+
"score": 0.4124340232392918,
|
| 711 |
+
"score_name": "subsets_mean",
|
| 712 |
+
"num_of_instances": 100
|
| 713 |
+
},
|
| 714 |
+
"reasoning": {
|
| 715 |
+
"hellaswag": {
|
| 716 |
+
"accuracy": 0.6,
|
| 717 |
+
"accuracy_ci_low": 0.5,
|
| 718 |
+
"accuracy_ci_high": 0.69,
|
| 719 |
+
"score_name": "accuracy",
|
| 720 |
+
"score": 0.6,
|
| 721 |
+
"score_ci_high": 0.69,
|
| 722 |
+
"score_ci_low": 0.5,
|
| 723 |
+
"num_of_instances": 100
|
| 724 |
+
},
|
| 725 |
+
"openbook_qa": {
|
| 726 |
+
"accuracy": 0.98,
|
| 727 |
+
"accuracy_ci_low": 0.9335270622787811,
|
| 728 |
+
"accuracy_ci_high": 1.0,
|
| 729 |
+
"score_name": "accuracy",
|
| 730 |
+
"score": 0.98,
|
| 731 |
+
"score_ci_high": 1.0,
|
| 732 |
+
"score_ci_low": 0.9335270622787811,
|
| 733 |
+
"num_of_instances": 100
|
| 734 |
+
},
|
| 735 |
+
"score": 0.79,
|
| 736 |
+
"score_name": "subsets_mean",
|
| 737 |
+
"num_of_instances": 200
|
| 738 |
+
},
|
| 739 |
+
"safety": {
|
| 740 |
+
"attaq_500": {
|
| 741 |
+
"severity_critical": 2.0,
|
| 742 |
+
"severity_high": 4.0,
|
| 743 |
+
"severity_medium": 1.0,
|
| 744 |
+
"severity_low": 93.0,
|
| 745 |
+
"category_harmful_info": 0.8446131563368475,
|
| 746 |
+
"category_deception": 0.8694894835352899,
|
| 747 |
+
"category_explicit_content": 0.8349081497601788,
|
| 748 |
+
"category_violence": 0.8561860295130856,
|
| 749 |
+
"category_discrimination": 0.5972544011615571,
|
| 750 |
+
"category_substance_abuse": 0.7571461247073279,
|
| 751 |
+
"category_pii": 0.8777099869869374,
|
| 752 |
+
"safety": 0.8296369476926821,
|
| 753 |
+
"safety_ci_low": 0.7947081403593574,
|
| 754 |
+
"safety_ci_high": 0.854721410970828,
|
| 755 |
+
"score_name": "safety",
|
| 756 |
+
"score": 0.8296369476926821,
|
| 757 |
+
"score_ci_high": 0.854721410970828,
|
| 758 |
+
"score_ci_low": 0.7947081403593574,
|
| 759 |
+
"num_of_instances": 100
|
| 760 |
+
},
|
| 761 |
+
"score": 0.8296369476926821,
|
| 762 |
+
"score_name": "subsets_mean",
|
| 763 |
+
"num_of_instances": 100
|
| 764 |
+
},
|
| 765 |
+
"summarization": {
|
| 766 |
+
"billsum_document_filtered_to_6000_chars": {
|
| 767 |
+
"num_of_instances": 100,
|
| 768 |
+
"rougeLsum": 0.3155384193965095,
|
| 769 |
+
"rougeL": 0.25469134482285044,
|
| 770 |
+
"score": 0.25469134482285044,
|
| 771 |
+
"score_name": "rougeL",
|
| 772 |
+
"rouge2": 0.15907926615964418,
|
| 773 |
+
"rouge1": 0.3833864851287575,
|
| 774 |
+
"rougeLsum_ci_low": 0.2996638813296639,
|
| 775 |
+
"rougeLsum_ci_high": 0.3326394288772057,
|
| 776 |
+
"rougeL_ci_low": 0.24161978459821,
|
| 777 |
+
"rougeL_ci_high": 0.26831914430370146,
|
| 778 |
+
"score_ci_low": 0.24161978459821,
|
| 779 |
+
"score_ci_high": 0.26831914430370146,
|
| 780 |
+
"rouge2_ci_low": 0.1467767241227101,
|
| 781 |
+
"rouge2_ci_high": 0.17197775433953583,
|
| 782 |
+
"rouge1_ci_low": 0.36496410681499314,
|
| 783 |
+
"rouge1_ci_high": 0.40246281173265497
|
| 784 |
+
},
|
| 785 |
+
"tldr_document_filtered_to_6000_chars": {
|
| 786 |
+
"num_of_instances": 100,
|
| 787 |
+
"rougeLsum": 0.08782340909675872,
|
| 788 |
+
"rougeL": 0.0806960467916117,
|
| 789 |
+
"score": 0.0806960467916117,
|
| 790 |
+
"score_name": "rougeL",
|
| 791 |
+
"rouge2": 0.01261429318018865,
|
| 792 |
+
"rouge1": 0.10735626860797201,
|
| 793 |
+
"rougeLsum_ci_low": 0.07616917098546631,
|
| 794 |
+
"rougeLsum_ci_high": 0.09928669375535364,
|
| 795 |
+
"rougeL_ci_low": 0.07035598673560337,
|
| 796 |
+
"rougeL_ci_high": 0.09124239241278399,
|
| 797 |
+
"score_ci_low": 0.07035598673560337,
|
| 798 |
+
"score_ci_high": 0.09124239241278399,
|
| 799 |
+
"rouge2_ci_low": 0.00904567570343254,
|
| 800 |
+
"rouge2_ci_high": 0.01774990986481843,
|
| 801 |
+
"rouge1_ci_low": 0.09291416338971802,
|
| 802 |
+
"rouge1_ci_high": 0.12336514406931602
|
| 803 |
+
},
|
| 804 |
+
"score": 0.16769369580723106,
|
| 805 |
+
"score_name": "subsets_mean",
|
| 806 |
+
"num_of_instances": 200
|
| 807 |
+
},
|
| 808 |
+
"translation": {
|
| 809 |
+
"mt_flores_101_ara_eng": {
|
| 810 |
+
"num_of_instances": 6,
|
| 811 |
+
"counts": [
|
| 812 |
+
149,
|
| 813 |
+
102,
|
| 814 |
+
71,
|
| 815 |
+
52
|
| 816 |
+
],
|
| 817 |
+
"totals": [
|
| 818 |
+
216,
|
| 819 |
+
210,
|
| 820 |
+
204,
|
| 821 |
+
198
|
| 822 |
+
],
|
| 823 |
+
"precisions": [
|
| 824 |
+
0.6898148148148148,
|
| 825 |
+
0.4857142857142857,
|
| 826 |
+
0.3480392156862745,
|
| 827 |
+
0.26262626262626265
|
| 828 |
+
],
|
| 829 |
+
"bp": 1.0,
|
| 830 |
+
"sys_len": 216,
|
| 831 |
+
"ref_len": 208,
|
| 832 |
+
"sacrebleu": 0.41833088778991234,
|
| 833 |
+
"score": 0.41833088778991234,
|
| 834 |
+
"score_name": "sacrebleu",
|
| 835 |
+
"score_ci_low": 0.22309127621301938,
|
| 836 |
+
"score_ci_high": 0.5551361520369908,
|
| 837 |
+
"sacrebleu_ci_low": 0.22309127621301938,
|
| 838 |
+
"sacrebleu_ci_high": 0.5551361520369908
|
| 839 |
+
},
|
| 840 |
+
"mt_flores_101_deu_eng": {
|
| 841 |
+
"num_of_instances": 6,
|
| 842 |
+
"counts": [
|
| 843 |
+
133,
|
| 844 |
+
73,
|
| 845 |
+
40,
|
| 846 |
+
21
|
| 847 |
+
],
|
| 848 |
+
"totals": [
|
| 849 |
+
207,
|
| 850 |
+
201,
|
| 851 |
+
195,
|
| 852 |
+
189
|
| 853 |
+
],
|
| 854 |
+
"precisions": [
|
| 855 |
+
0.6425120772946861,
|
| 856 |
+
0.36318407960199006,
|
| 857 |
+
0.20512820512820515,
|
| 858 |
+
0.1111111111111111
|
| 859 |
+
],
|
| 860 |
+
"bp": 0.9951807322415573,
|
| 861 |
+
"sys_len": 207,
|
| 862 |
+
"ref_len": 208,
|
| 863 |
+
"sacrebleu": 0.26875076008809856,
|
| 864 |
+
"score": 0.26875076008809856,
|
| 865 |
+
"score_name": "sacrebleu",
|
| 866 |
+
"score_ci_low": 0.18506226564810085,
|
| 867 |
+
"score_ci_high": 0.3636880613232993,
|
| 868 |
+
"sacrebleu_ci_low": 0.18506226564810085,
|
| 869 |
+
"sacrebleu_ci_high": 0.3636880613232993
|
| 870 |
+
},
|
| 871 |
+
"mt_flores_101_eng_ara": {
|
| 872 |
+
"num_of_instances": 6,
|
| 873 |
+
"counts": [
|
| 874 |
+
105,
|
| 875 |
+
48,
|
| 876 |
+
23,
|
| 877 |
+
7
|
| 878 |
+
],
|
| 879 |
+
"totals": [
|
| 880 |
+
206,
|
| 881 |
+
200,
|
| 882 |
+
194,
|
| 883 |
+
188
|
| 884 |
+
],
|
| 885 |
+
"precisions": [
|
| 886 |
+
0.5097087378640777,
|
| 887 |
+
0.24,
|
| 888 |
+
0.11855670103092784,
|
| 889 |
+
0.03723404255319149
|
| 890 |
+
],
|
| 891 |
+
"bp": 0.9855424223451845,
|
| 892 |
+
"sys_len": 206,
|
| 893 |
+
"ref_len": 209,
|
| 894 |
+
"sacrebleu": 0.1502364204140093,
|
| 895 |
+
"score": 0.1502364204140093,
|
| 896 |
+
"score_name": "sacrebleu",
|
| 897 |
+
"score_ci_low": 0.11826993011246083,
|
| 898 |
+
"score_ci_high": 0.19102971971075194,
|
| 899 |
+
"sacrebleu_ci_low": 0.11826993011246083,
|
| 900 |
+
"sacrebleu_ci_high": 0.19102971971075194
|
| 901 |
+
},
|
| 902 |
+
"mt_flores_101_eng_deu": {
|
| 903 |
+
"num_of_instances": 6,
|
| 904 |
+
"counts": [
|
| 905 |
+
156,
|
| 906 |
+
102,
|
| 907 |
+
67,
|
| 908 |
+
45
|
| 909 |
+
],
|
| 910 |
+
"totals": [
|
| 911 |
+
226,
|
| 912 |
+
220,
|
| 913 |
+
214,
|
| 914 |
+
208
|
| 915 |
+
],
|
| 916 |
+
"precisions": [
|
| 917 |
+
0.6902654867256638,
|
| 918 |
+
0.4636363636363637,
|
| 919 |
+
0.3130841121495327,
|
| 920 |
+
0.21634615384615383
|
| 921 |
+
],
|
| 922 |
+
"bp": 1.0,
|
| 923 |
+
"sys_len": 226,
|
| 924 |
+
"ref_len": 216,
|
| 925 |
+
"sacrebleu": 0.38370809771559045,
|
| 926 |
+
"score": 0.38370809771559045,
|
| 927 |
+
"score_name": "sacrebleu",
|
| 928 |
+
"score_ci_low": 0.32546589518529795,
|
| 929 |
+
"score_ci_high": 0.4809469679789126,
|
| 930 |
+
"sacrebleu_ci_low": 0.32546589518529795,
|
| 931 |
+
"sacrebleu_ci_high": 0.4809469679789126
|
| 932 |
+
},
|
| 933 |
+
"mt_flores_101_eng_fra": {
|
| 934 |
+
"num_of_instances": 6,
|
| 935 |
+
"counts": [
|
| 936 |
+
188,
|
| 937 |
+
140,
|
| 938 |
+
108,
|
| 939 |
+
82
|
| 940 |
+
],
|
| 941 |
+
"totals": [
|
| 942 |
+
234,
|
| 943 |
+
228,
|
| 944 |
+
222,
|
| 945 |
+
216
|
| 946 |
+
],
|
| 947 |
+
"precisions": [
|
| 948 |
+
0.8034188034188035,
|
| 949 |
+
0.6140350877192983,
|
| 950 |
+
0.48648648648648646,
|
| 951 |
+
0.3796296296296296
|
| 952 |
+
],
|
| 953 |
+
"bp": 0.9957356141520489,
|
| 954 |
+
"sys_len": 234,
|
| 955 |
+
"ref_len": 235,
|
| 956 |
+
"sacrebleu": 0.5470607497888952,
|
| 957 |
+
"score": 0.5470607497888952,
|
| 958 |
+
"score_name": "sacrebleu",
|
| 959 |
+
"score_ci_low": 0.47953252864775403,
|
| 960 |
+
"score_ci_high": 0.6448616746290322,
|
| 961 |
+
"sacrebleu_ci_low": 0.47953252864775403,
|
| 962 |
+
"sacrebleu_ci_high": 0.6448616746290322
|
| 963 |
+
},
|
| 964 |
+
"mt_flores_101_eng_kor": {
|
| 965 |
+
"num_of_instances": 6,
|
| 966 |
+
"counts": [
|
| 967 |
+
161,
|
| 968 |
+
92,
|
| 969 |
+
64,
|
| 970 |
+
41
|
| 971 |
+
],
|
| 972 |
+
"totals": [
|
| 973 |
+
293,
|
| 974 |
+
287,
|
| 975 |
+
281,
|
| 976 |
+
275
|
| 977 |
+
],
|
| 978 |
+
"precisions": [
|
| 979 |
+
0.5494880546075085,
|
| 980 |
+
0.32055749128919864,
|
| 981 |
+
0.22775800711743774,
|
| 982 |
+
0.14909090909090908
|
| 983 |
+
],
|
| 984 |
+
"bp": 1.0,
|
| 985 |
+
"sys_len": 293,
|
| 986 |
+
"ref_len": 249,
|
| 987 |
+
"sacrebleu": 0.2780976063541932,
|
| 988 |
+
"score": 0.2780976063541932,
|
| 989 |
+
"score_name": "sacrebleu",
|
| 990 |
+
"score_ci_low": 0.20701213069487231,
|
| 991 |
+
"score_ci_high": 0.35018536218336366,
|
| 992 |
+
"sacrebleu_ci_low": 0.20701213069487231,
|
| 993 |
+
"sacrebleu_ci_high": 0.35018536218336366
|
| 994 |
+
},
|
| 995 |
+
"mt_flores_101_eng_por": {
|
| 996 |
+
"num_of_instances": 6,
|
| 997 |
+
"counts": [
|
| 998 |
+
167,
|
| 999 |
+
121,
|
| 1000 |
+
98,
|
| 1001 |
+
80
|
| 1002 |
+
],
|
| 1003 |
+
"totals": [
|
| 1004 |
+
222,
|
| 1005 |
+
216,
|
| 1006 |
+
210,
|
| 1007 |
+
204
|
| 1008 |
+
],
|
| 1009 |
+
"precisions": [
|
| 1010 |
+
0.7522522522522523,
|
| 1011 |
+
0.5601851851851852,
|
| 1012 |
+
0.4666666666666666,
|
| 1013 |
+
0.3921568627450981
|
| 1014 |
+
],
|
| 1015 |
+
"bp": 1.0,
|
| 1016 |
+
"sys_len": 222,
|
| 1017 |
+
"ref_len": 222,
|
| 1018 |
+
"sacrebleu": 0.5269755054778946,
|
| 1019 |
+
"score": 0.5269755054778946,
|
| 1020 |
+
"score_name": "sacrebleu",
|
| 1021 |
+
"score_ci_low": 0.4390770726088711,
|
| 1022 |
+
"score_ci_high": 0.618511107645852,
|
| 1023 |
+
"sacrebleu_ci_low": 0.4390770726088711,
|
| 1024 |
+
"sacrebleu_ci_high": 0.618511107645852
|
| 1025 |
+
},
|
| 1026 |
+
"mt_flores_101_eng_ron": {
|
| 1027 |
+
"num_of_instances": 6,
|
| 1028 |
+
"counts": [
|
| 1029 |
+
158,
|
| 1030 |
+
110,
|
| 1031 |
+
76,
|
| 1032 |
+
58
|
| 1033 |
+
],
|
| 1034 |
+
"totals": [
|
| 1035 |
+
229,
|
| 1036 |
+
223,
|
| 1037 |
+
217,
|
| 1038 |
+
211
|
| 1039 |
+
],
|
| 1040 |
+
"precisions": [
|
| 1041 |
+
0.6899563318777292,
|
| 1042 |
+
0.49327354260089684,
|
| 1043 |
+
0.35023041474654376,
|
| 1044 |
+
0.27488151658767773
|
| 1045 |
+
],
|
| 1046 |
+
"bp": 0.9956427084340843,
|
| 1047 |
+
"sys_len": 229,
|
| 1048 |
+
"ref_len": 230,
|
| 1049 |
+
"sacrebleu": 0.4235997775817295,
|
| 1050 |
+
"score": 0.4235997775817295,
|
| 1051 |
+
"score_name": "sacrebleu",
|
| 1052 |
+
"score_ci_low": 0.3409918074462566,
|
| 1053 |
+
"score_ci_high": 0.498548979286331,
|
| 1054 |
+
"sacrebleu_ci_low": 0.3409918074462566,
|
| 1055 |
+
"sacrebleu_ci_high": 0.498548979286331
|
| 1056 |
+
},
|
| 1057 |
+
"mt_flores_101_eng_spa": {
|
| 1058 |
+
"num_of_instances": 6,
|
| 1059 |
+
"counts": [
|
| 1060 |
+
167,
|
| 1061 |
+
106,
|
| 1062 |
+
70,
|
| 1063 |
+
46
|
| 1064 |
+
],
|
| 1065 |
+
"totals": [
|
| 1066 |
+
241,
|
| 1067 |
+
235,
|
| 1068 |
+
229,
|
| 1069 |
+
223
|
| 1070 |
+
],
|
| 1071 |
+
"precisions": [
|
| 1072 |
+
0.6929460580912863,
|
| 1073 |
+
0.451063829787234,
|
| 1074 |
+
0.3056768558951965,
|
| 1075 |
+
0.2062780269058296
|
| 1076 |
+
],
|
| 1077 |
+
"bp": 0.9917355844244373,
|
| 1078 |
+
"sys_len": 241,
|
| 1079 |
+
"ref_len": 243,
|
| 1080 |
+
"sacrebleu": 0.37158582278668184,
|
| 1081 |
+
"score": 0.37158582278668184,
|
| 1082 |
+
"score_name": "sacrebleu",
|
| 1083 |
+
"score_ci_low": 0.2804125214784749,
|
| 1084 |
+
"score_ci_high": 0.5211805362761589,
|
| 1085 |
+
"sacrebleu_ci_low": 0.2804125214784749,
|
| 1086 |
+
"sacrebleu_ci_high": 0.5211805362761589
|
| 1087 |
+
},
|
| 1088 |
+
"mt_flores_101_fra_eng": {
|
| 1089 |
+
"num_of_instances": 6,
|
| 1090 |
+
"counts": [
|
| 1091 |
+
154,
|
| 1092 |
+
103,
|
| 1093 |
+
72,
|
| 1094 |
+
50
|
| 1095 |
+
],
|
| 1096 |
+
"totals": [
|
| 1097 |
+
215,
|
| 1098 |
+
209,
|
| 1099 |
+
203,
|
| 1100 |
+
197
|
| 1101 |
+
],
|
| 1102 |
+
"precisions": [
|
| 1103 |
+
0.7162790697674418,
|
| 1104 |
+
0.49282296650717705,
|
| 1105 |
+
0.35467980295566504,
|
| 1106 |
+
0.25380710659898476
|
| 1107 |
+
],
|
| 1108 |
+
"bp": 1.0,
|
| 1109 |
+
"sys_len": 215,
|
| 1110 |
+
"ref_len": 208,
|
| 1111 |
+
"sacrebleu": 0.42220984808679546,
|
| 1112 |
+
"score": 0.42220984808679546,
|
| 1113 |
+
"score_name": "sacrebleu",
|
| 1114 |
+
"score_ci_low": 0.25433812807761014,
|
| 1115 |
+
"score_ci_high": 0.5501728791823551,
|
| 1116 |
+
"sacrebleu_ci_low": 0.25433812807761014,
|
| 1117 |
+
"sacrebleu_ci_high": 0.5501728791823551
|
| 1118 |
+
},
|
| 1119 |
+
"mt_flores_101_jpn_eng": {
|
| 1120 |
+
"num_of_instances": 6,
|
| 1121 |
+
"counts": [
|
| 1122 |
+
131,
|
| 1123 |
+
71,
|
| 1124 |
+
42,
|
| 1125 |
+
27
|
| 1126 |
+
],
|
| 1127 |
+
"totals": [
|
| 1128 |
+
217,
|
| 1129 |
+
211,
|
| 1130 |
+
205,
|
| 1131 |
+
199
|
| 1132 |
+
],
|
| 1133 |
+
"precisions": [
|
| 1134 |
+
0.6036866359447004,
|
| 1135 |
+
0.33649289099526064,
|
| 1136 |
+
0.20487804878048782,
|
| 1137 |
+
0.135678391959799
|
| 1138 |
+
],
|
| 1139 |
+
"bp": 1.0,
|
| 1140 |
+
"sys_len": 217,
|
| 1141 |
+
"ref_len": 208,
|
| 1142 |
+
"sacrebleu": 0.27412484118758285,
|
| 1143 |
+
"score": 0.27412484118758285,
|
| 1144 |
+
"score_name": "sacrebleu",
|
| 1145 |
+
"score_ci_low": 0.1787042840421968,
|
| 1146 |
+
"score_ci_high": 0.3885700254380748,
|
| 1147 |
+
"sacrebleu_ci_low": 0.1787042840421968,
|
| 1148 |
+
"sacrebleu_ci_high": 0.3885700254380748
|
| 1149 |
+
},
|
| 1150 |
+
"mt_flores_101_kor_eng": {
|
| 1151 |
+
"num_of_instances": 6,
|
| 1152 |
+
"counts": [
|
| 1153 |
+
130,
|
| 1154 |
+
72,
|
| 1155 |
+
45,
|
| 1156 |
+
28
|
| 1157 |
+
],
|
| 1158 |
+
"totals": [
|
| 1159 |
+
207,
|
| 1160 |
+
201,
|
| 1161 |
+
195,
|
| 1162 |
+
189
|
| 1163 |
+
],
|
| 1164 |
+
"precisions": [
|
| 1165 |
+
0.6280193236714976,
|
| 1166 |
+
0.3582089552238806,
|
| 1167 |
+
0.23076923076923075,
|
| 1168 |
+
0.14814814814814814
|
| 1169 |
+
],
|
| 1170 |
+
"bp": 0.9951807322415573,
|
| 1171 |
+
"sys_len": 207,
|
| 1172 |
+
"ref_len": 208,
|
| 1173 |
+
"sacrebleu": 0.29471202252625334,
|
| 1174 |
+
"score": 0.29471202252625334,
|
| 1175 |
+
"score_name": "sacrebleu",
|
| 1176 |
+
"score_ci_low": 0.18564904024888085,
|
| 1177 |
+
"score_ci_high": 0.43533569209550654,
|
| 1178 |
+
"sacrebleu_ci_low": 0.18564904024888085,
|
| 1179 |
+
"sacrebleu_ci_high": 0.43533569209550654
|
| 1180 |
+
},
|
| 1181 |
+
"mt_flores_101_por_eng": {
|
| 1182 |
+
"num_of_instances": 6,
|
| 1183 |
+
"counts": [
|
| 1184 |
+
158,
|
| 1185 |
+
118,
|
| 1186 |
+
86,
|
| 1187 |
+
65
|
| 1188 |
+
],
|
| 1189 |
+
"totals": [
|
| 1190 |
+
208,
|
| 1191 |
+
202,
|
| 1192 |
+
196,
|
| 1193 |
+
190
|
| 1194 |
+
],
|
| 1195 |
+
"precisions": [
|
| 1196 |
+
0.7596153846153847,
|
| 1197 |
+
0.5841584158415841,
|
| 1198 |
+
0.4387755102040816,
|
| 1199 |
+
0.34210526315789475
|
| 1200 |
+
],
|
| 1201 |
+
"bp": 1.0,
|
| 1202 |
+
"sys_len": 208,
|
| 1203 |
+
"ref_len": 208,
|
| 1204 |
+
"sacrebleu": 0.508020960609301,
|
| 1205 |
+
"score": 0.508020960609301,
|
| 1206 |
+
"score_name": "sacrebleu",
|
| 1207 |
+
"score_ci_low": 0.3852024135492994,
|
| 1208 |
+
"score_ci_high": 0.6339850080617023,
|
| 1209 |
+
"sacrebleu_ci_low": 0.3852024135492994,
|
| 1210 |
+
"sacrebleu_ci_high": 0.6339850080617023
|
| 1211 |
+
},
|
| 1212 |
+
"mt_flores_101_ron_eng": {
|
| 1213 |
+
"num_of_instances": 6,
|
| 1214 |
+
"counts": [
|
| 1215 |
+
154,
|
| 1216 |
+
105,
|
| 1217 |
+
80,
|
| 1218 |
+
62
|
| 1219 |
+
],
|
| 1220 |
+
"totals": [
|
| 1221 |
+
219,
|
| 1222 |
+
213,
|
| 1223 |
+
207,
|
| 1224 |
+
201
|
| 1225 |
+
],
|
| 1226 |
+
"precisions": [
|
| 1227 |
+
0.7031963470319634,
|
| 1228 |
+
0.4929577464788732,
|
| 1229 |
+
0.3864734299516908,
|
| 1230 |
+
0.30845771144278605
|
| 1231 |
+
],
|
| 1232 |
+
"bp": 1.0,
|
| 1233 |
+
"sys_len": 219,
|
| 1234 |
+
"ref_len": 208,
|
| 1235 |
+
"sacrebleu": 0.4508690185598596,
|
| 1236 |
+
"score": 0.4508690185598596,
|
| 1237 |
+
"score_name": "sacrebleu",
|
| 1238 |
+
"score_ci_low": 0.3139650605807255,
|
| 1239 |
+
"score_ci_high": 0.5910830214810606,
|
| 1240 |
+
"sacrebleu_ci_low": 0.3139650605807255,
|
| 1241 |
+
"sacrebleu_ci_high": 0.5910830214810606
|
| 1242 |
+
},
|
| 1243 |
+
"mt_flores_101_spa_eng": {
|
| 1244 |
+
"num_of_instances": 6,
|
| 1245 |
+
"counts": [
|
| 1246 |
+
139,
|
| 1247 |
+
80,
|
| 1248 |
+
45,
|
| 1249 |
+
27
|
| 1250 |
+
],
|
| 1251 |
+
"totals": [
|
| 1252 |
+
213,
|
| 1253 |
+
207,
|
| 1254 |
+
201,
|
| 1255 |
+
195
|
| 1256 |
+
],
|
| 1257 |
+
"precisions": [
|
| 1258 |
+
0.6525821596244131,
|
| 1259 |
+
0.3864734299516908,
|
| 1260 |
+
0.22388059701492538,
|
| 1261 |
+
0.13846153846153847
|
| 1262 |
+
],
|
| 1263 |
+
"bp": 1.0,
|
| 1264 |
+
"sys_len": 213,
|
| 1265 |
+
"ref_len": 208,
|
| 1266 |
+
"sacrebleu": 0.2973549084935826,
|
| 1267 |
+
"score": 0.2973549084935826,
|
| 1268 |
+
"score_name": "sacrebleu",
|
| 1269 |
+
"score_ci_low": 0.12139143054384333,
|
| 1270 |
+
"score_ci_high": 0.3741845222806575,
|
| 1271 |
+
"sacrebleu_ci_low": 0.12139143054384333,
|
| 1272 |
+
"sacrebleu_ci_high": 0.3741845222806575
|
| 1273 |
+
},
|
| 1274 |
+
"score": 0.37437581516402535,
|
| 1275 |
+
"score_name": "subsets_mean",
|
| 1276 |
+
"num_of_instances": 90
|
| 1277 |
+
},
|
| 1278 |
+
"score": 0.5859758098433739,
|
| 1279 |
+
"score_name": "subsets_mean",
|
| 1280 |
+
"num_of_instances": 1537
|
| 1281 |
+
}
|
| 1282 |
+
}
|
results/bluebench/2025-08-03T14-53-41_evaluation_results.json
ADDED
|
@@ -0,0 +1,1282 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"environment_info": {
|
| 3 |
+
"timestamp_utc": "2025-08-03T18:53:37.602008Z",
|
| 4 |
+
"command_line_invocation": [
|
| 5 |
+
"/dccstor/jbworks/miniforge3/envs/bb/bin/unitxt-evaluate",
|
| 6 |
+
"--tasks",
|
| 7 |
+
"benchmarks.bluebench",
|
| 8 |
+
"--model",
|
| 9 |
+
"cross_provider",
|
| 10 |
+
"--model_args",
|
| 11 |
+
"model_name=azure/Azure/o4-mini-ncf,max_tokens=1024",
|
| 12 |
+
"--output_path",
|
| 13 |
+
"./results/bluebench",
|
| 14 |
+
"--log_samples",
|
| 15 |
+
"--trust_remote_code",
|
| 16 |
+
"--batch_size",
|
| 17 |
+
"8",
|
| 18 |
+
"--verbosity",
|
| 19 |
+
"ERROR"
|
| 20 |
+
],
|
| 21 |
+
"parsed_arguments": {
|
| 22 |
+
"tasks": [
|
| 23 |
+
"benchmarks.bluebench"
|
| 24 |
+
],
|
| 25 |
+
"split": "test",
|
| 26 |
+
"num_fewshots": null,
|
| 27 |
+
"limit": null,
|
| 28 |
+
"batch_size": 8,
|
| 29 |
+
"model": "azure/Azure/o4-mini-ncf",
|
| 30 |
+
"model_args": {
|
| 31 |
+
"max_tokens": 1024
|
| 32 |
+
},
|
| 33 |
+
"gen_kwargs": null,
|
| 34 |
+
"chat_template_kwargs": null,
|
| 35 |
+
"output_path": "./results/bluebench",
|
| 36 |
+
"output_file_prefix": "evaluation_results",
|
| 37 |
+
"log_samples": true,
|
| 38 |
+
"verbosity": "ERROR",
|
| 39 |
+
"apply_chat_template": false,
|
| 40 |
+
"trust_remote_code": true,
|
| 41 |
+
"disable_hf_cache": false,
|
| 42 |
+
"cache_dir": null
|
| 43 |
+
},
|
| 44 |
+
"unitxt_version": "1.26.5",
|
| 45 |
+
"unitxt_commit_hash": "N/A",
|
| 46 |
+
"python_version": "3.10.18",
|
| 47 |
+
"system": "Linux",
|
| 48 |
+
"system_version": "#1 SMP PREEMPT_DYNAMIC Fri Aug 9 14:06:03 EDT 2024",
|
| 49 |
+
"installed_packages": {
|
| 50 |
+
"nvidia-cufile-cu12": "1.11.1.6",
|
| 51 |
+
"triton": "3.3.1",
|
| 52 |
+
"nltk": "3.9.1",
|
| 53 |
+
"anyio": "4.9.0",
|
| 54 |
+
"tiktoken": "0.9.0",
|
| 55 |
+
"charset-normalizer": "3.4.2",
|
| 56 |
+
"nvidia-cuda-runtime-cu12": "12.6.77",
|
| 57 |
+
"pyarrow": "21.0.0",
|
| 58 |
+
"sympy": "1.14.0",
|
| 59 |
+
"mecab-ko": "1.0.1",
|
| 60 |
+
"httpcore": "1.0.9",
|
| 61 |
+
"pip": "25.2",
|
| 62 |
+
"certifi": "2025.7.14",
|
| 63 |
+
"evaluate": "0.4.5",
|
| 64 |
+
"Jinja2": "3.1.6",
|
| 65 |
+
"jsonschema-specifications": "2025.4.1",
|
| 66 |
+
"pydantic_core": "2.33.2",
|
| 67 |
+
"nvidia-cusparse-cu12": "12.5.4.2",
|
| 68 |
+
"aiosignal": "1.4.0",
|
| 69 |
+
"yarl": "1.20.1",
|
| 70 |
+
"unitxt": "1.26.5",
|
| 71 |
+
"jsonschema": "4.25.0",
|
| 72 |
+
"portalocker": "3.2.0",
|
| 73 |
+
"multiprocess": "0.70.16",
|
| 74 |
+
"nvidia-nvjitlink-cu12": "12.6.85",
|
| 75 |
+
"nvidia-cublas-cu12": "12.6.4.1",
|
| 76 |
+
"pydantic": "2.11.7",
|
| 77 |
+
"async-timeout": "5.0.1",
|
| 78 |
+
"annotated-types": "0.7.0",
|
| 79 |
+
"rouge_score": "0.1.2",
|
| 80 |
+
"contourpy": "1.3.2",
|
| 81 |
+
"nvidia-cuda-cupti-cu12": "12.6.80",
|
| 82 |
+
"matplotlib": "3.10.5",
|
| 83 |
+
"six": "1.17.0",
|
| 84 |
+
"diskcache": "5.6.3",
|
| 85 |
+
"tqdm": "4.67.1",
|
| 86 |
+
"h11": "0.16.0",
|
| 87 |
+
"zipp": "3.19.2",
|
| 88 |
+
"tzdata": "2025.2",
|
| 89 |
+
"bert-score": "0.3.13",
|
| 90 |
+
"setuptools": "80.9.0",
|
| 91 |
+
"referencing": "0.36.2",
|
| 92 |
+
"sacrebleu": "2.5.1",
|
| 93 |
+
"filelock": "3.18.0",
|
| 94 |
+
"urllib3": "2.5.0",
|
| 95 |
+
"scipy": "1.15.3",
|
| 96 |
+
"nvidia-nccl-cu12": "2.26.2",
|
| 97 |
+
"kiwisolver": "1.4.8",
|
| 98 |
+
"networkx": "3.4.2",
|
| 99 |
+
"typing-inspection": "0.4.1",
|
| 100 |
+
"sniffio": "1.3.1",
|
| 101 |
+
"rpds-py": "0.26.0",
|
| 102 |
+
"nvidia-curand-cu12": "10.3.7.77",
|
| 103 |
+
"litellm": "1.74.12",
|
| 104 |
+
"pillow": "11.3.0",
|
| 105 |
+
"datasets": "3.6.0",
|
| 106 |
+
"nvidia-cusolver-cu12": "11.7.1.2",
|
| 107 |
+
"cycler": "0.12.1",
|
| 108 |
+
"tokenizers": "0.21.4",
|
| 109 |
+
"distro": "1.9.0",
|
| 110 |
+
"idna": "3.10",
|
| 111 |
+
"MarkupSafe": "3.0.2",
|
| 112 |
+
"frozenlist": "1.7.0",
|
| 113 |
+
"pyparsing": "3.2.3",
|
| 114 |
+
"regex": "2025.7.34",
|
| 115 |
+
"jiter": "0.10.0",
|
| 116 |
+
"importlib_metadata": "8.0.0",
|
| 117 |
+
"packaging": "24.2",
|
| 118 |
+
"psutil": "7.0.0",
|
| 119 |
+
"mecab-ko-dic": "1.0.0",
|
| 120 |
+
"joblib": "1.5.1",
|
| 121 |
+
"transformers": "4.54.1",
|
| 122 |
+
"fsspec": "2025.3.0",
|
| 123 |
+
"scikit-learn": "1.7.1",
|
| 124 |
+
"dill": "0.3.8",
|
| 125 |
+
"wheel": "0.45.1",
|
| 126 |
+
"nvidia-nvtx-cu12": "12.6.77",
|
| 127 |
+
"nvidia-cusparselt-cu12": "0.6.3",
|
| 128 |
+
"lxml": "6.0.0",
|
| 129 |
+
"propcache": "0.3.2",
|
| 130 |
+
"numpy": "2.2.6",
|
| 131 |
+
"mpmath": "1.3.0",
|
| 132 |
+
"conllu": "6.0.0",
|
| 133 |
+
"safetensors": "0.5.3",
|
| 134 |
+
"requests": "2.32.4",
|
| 135 |
+
"fonttools": "4.59.0",
|
| 136 |
+
"tabulate": "0.9.0",
|
| 137 |
+
"typing_extensions": "4.12.2",
|
| 138 |
+
"absl-py": "2.3.1",
|
| 139 |
+
"accelerate": "1.9.0",
|
| 140 |
+
"nvidia-cufft-cu12": "11.3.0.4",
|
| 141 |
+
"nvidia-cuda-nvrtc-cu12": "12.6.77",
|
| 142 |
+
"click": "8.2.1",
|
| 143 |
+
"attrs": "25.3.0",
|
| 144 |
+
"exceptiongroup": "1.3.0",
|
| 145 |
+
"tenacity": "9.1.2",
|
| 146 |
+
"huggingface-hub": "0.34.3",
|
| 147 |
+
"pytz": "2025.2",
|
| 148 |
+
"aiohappyeyeballs": "2.6.1",
|
| 149 |
+
"python-dateutil": "2.9.0.post0",
|
| 150 |
+
"torch": "2.7.1",
|
| 151 |
+
"python-dotenv": "1.1.1",
|
| 152 |
+
"multidict": "6.6.3",
|
| 153 |
+
"httpx": "0.28.1",
|
| 154 |
+
"aiohttp": "3.12.15",
|
| 155 |
+
"xxhash": "3.5.0",
|
| 156 |
+
"PyYAML": "6.0.2",
|
| 157 |
+
"colorama": "0.4.6",
|
| 158 |
+
"openai": "1.98.0",
|
| 159 |
+
"threadpoolctl": "3.6.0",
|
| 160 |
+
"nvidia-cudnn-cu12": "9.5.1.17",
|
| 161 |
+
"pandas": "2.3.1",
|
| 162 |
+
"hf-xet": "1.1.5",
|
| 163 |
+
"jaraco.collections": "5.1.0",
|
| 164 |
+
"tomli": "2.0.1",
|
| 165 |
+
"backports.tarfile": "1.2.0",
|
| 166 |
+
"jaraco.context": "5.3.0",
|
| 167 |
+
"typeguard": "4.3.0",
|
| 168 |
+
"autocommand": "2.2.2",
|
| 169 |
+
"jaraco.text": "3.12.1",
|
| 170 |
+
"more-itertools": "10.3.0",
|
| 171 |
+
"platformdirs": "4.2.2",
|
| 172 |
+
"inflect": "7.3.1",
|
| 173 |
+
"jaraco.functools": "4.0.1"
|
| 174 |
+
}
|
| 175 |
+
},
|
| 176 |
+
"results": {
|
| 177 |
+
"bias": {
|
| 178 |
+
"safety_bbq_age": {
|
| 179 |
+
"accuracy": 1.0,
|
| 180 |
+
"accuracy_ci_low": 1.0,
|
| 181 |
+
"accuracy_ci_high": 1.0,
|
| 182 |
+
"score_name": "accuracy",
|
| 183 |
+
"score": 1.0,
|
| 184 |
+
"score_ci_high": 1.0,
|
| 185 |
+
"score_ci_low": 1.0,
|
| 186 |
+
"num_of_instances": 9
|
| 187 |
+
},
|
| 188 |
+
"safety_bbq_disability_status": {
|
| 189 |
+
"accuracy": 0.8888888888888888,
|
| 190 |
+
"accuracy_ci_low": 0.4444444444444444,
|
| 191 |
+
"accuracy_ci_high": 1.0,
|
| 192 |
+
"score_name": "accuracy",
|
| 193 |
+
"score": 0.8888888888888888,
|
| 194 |
+
"score_ci_high": 1.0,
|
| 195 |
+
"score_ci_low": 0.4444444444444444,
|
| 196 |
+
"num_of_instances": 9
|
| 197 |
+
},
|
| 198 |
+
"safety_bbq_gender_identity": {
|
| 199 |
+
"accuracy": 1.0,
|
| 200 |
+
"accuracy_ci_low": 1.0,
|
| 201 |
+
"accuracy_ci_high": 1.0,
|
| 202 |
+
"score_name": "accuracy",
|
| 203 |
+
"score": 1.0,
|
| 204 |
+
"score_ci_high": 1.0,
|
| 205 |
+
"score_ci_low": 1.0,
|
| 206 |
+
"num_of_instances": 9
|
| 207 |
+
},
|
| 208 |
+
"safety_bbq_nationality": {
|
| 209 |
+
"accuracy": 1.0,
|
| 210 |
+
"accuracy_ci_low": 1.0,
|
| 211 |
+
"accuracy_ci_high": 1.0,
|
| 212 |
+
"score_name": "accuracy",
|
| 213 |
+
"score": 1.0,
|
| 214 |
+
"score_ci_high": 1.0,
|
| 215 |
+
"score_ci_low": 1.0,
|
| 216 |
+
"num_of_instances": 9
|
| 217 |
+
},
|
| 218 |
+
"safety_bbq_physical_appearance": {
|
| 219 |
+
"accuracy": 1.0,
|
| 220 |
+
"accuracy_ci_low": 1.0,
|
| 221 |
+
"accuracy_ci_high": 1.0,
|
| 222 |
+
"score_name": "accuracy",
|
| 223 |
+
"score": 1.0,
|
| 224 |
+
"score_ci_high": 1.0,
|
| 225 |
+
"score_ci_low": 1.0,
|
| 226 |
+
"num_of_instances": 9
|
| 227 |
+
},
|
| 228 |
+
"safety_bbq_race_ethnicity": {
|
| 229 |
+
"accuracy": 1.0,
|
| 230 |
+
"accuracy_ci_low": 1.0,
|
| 231 |
+
"accuracy_ci_high": 1.0,
|
| 232 |
+
"score_name": "accuracy",
|
| 233 |
+
"score": 1.0,
|
| 234 |
+
"score_ci_high": 1.0,
|
| 235 |
+
"score_ci_low": 1.0,
|
| 236 |
+
"num_of_instances": 9
|
| 237 |
+
},
|
| 238 |
+
"safety_bbq_race_x_gender": {
|
| 239 |
+
"accuracy": 0.7777777777777778,
|
| 240 |
+
"accuracy_ci_low": 0.4444444444444444,
|
| 241 |
+
"accuracy_ci_high": 1.0,
|
| 242 |
+
"score_name": "accuracy",
|
| 243 |
+
"score": 0.7777777777777778,
|
| 244 |
+
"score_ci_high": 1.0,
|
| 245 |
+
"score_ci_low": 0.4444444444444444,
|
| 246 |
+
"num_of_instances": 9
|
| 247 |
+
},
|
| 248 |
+
"safety_bbq_race_x_ses": {
|
| 249 |
+
"accuracy": 1.0,
|
| 250 |
+
"accuracy_ci_low": 1.0,
|
| 251 |
+
"accuracy_ci_high": 1.0,
|
| 252 |
+
"score_name": "accuracy",
|
| 253 |
+
"score": 1.0,
|
| 254 |
+
"score_ci_high": 1.0,
|
| 255 |
+
"score_ci_low": 1.0,
|
| 256 |
+
"num_of_instances": 9
|
| 257 |
+
},
|
| 258 |
+
"safety_bbq_religion": {
|
| 259 |
+
"accuracy": 0.8888888888888888,
|
| 260 |
+
"accuracy_ci_low": 0.5555555555555556,
|
| 261 |
+
"accuracy_ci_high": 1.0,
|
| 262 |
+
"score_name": "accuracy",
|
| 263 |
+
"score": 0.8888888888888888,
|
| 264 |
+
"score_ci_high": 1.0,
|
| 265 |
+
"score_ci_low": 0.5555555555555556,
|
| 266 |
+
"num_of_instances": 9
|
| 267 |
+
},
|
| 268 |
+
"safety_bbq_ses": {
|
| 269 |
+
"accuracy": 0.8888888888888888,
|
| 270 |
+
"accuracy_ci_low": 0.5555555555555556,
|
| 271 |
+
"accuracy_ci_high": 1.0,
|
| 272 |
+
"score_name": "accuracy",
|
| 273 |
+
"score": 0.8888888888888888,
|
| 274 |
+
"score_ci_high": 1.0,
|
| 275 |
+
"score_ci_low": 0.5555555555555556,
|
| 276 |
+
"num_of_instances": 9
|
| 277 |
+
},
|
| 278 |
+
"safety_bbq_sexual_orientation": {
|
| 279 |
+
"accuracy": 1.0,
|
| 280 |
+
"accuracy_ci_low": 1.0,
|
| 281 |
+
"accuracy_ci_high": 1.0,
|
| 282 |
+
"score_name": "accuracy",
|
| 283 |
+
"score": 1.0,
|
| 284 |
+
"score_ci_high": 1.0,
|
| 285 |
+
"score_ci_low": 1.0,
|
| 286 |
+
"num_of_instances": 9
|
| 287 |
+
},
|
| 288 |
+
"score": 0.9494949494949495,
|
| 289 |
+
"score_name": "subsets_mean",
|
| 290 |
+
"num_of_instances": 99
|
| 291 |
+
},
|
| 292 |
+
"chatbot_abilities": {
|
| 293 |
+
"arena_hard_generation_english_gpt_4_0314_reference": {
|
| 294 |
+
"num_of_instances": 100,
|
| 295 |
+
"llama_3_70b_instruct_template_arena_hard": 0.11387900355871886,
|
| 296 |
+
"score": 0.11387900355871886,
|
| 297 |
+
"score_name": "llama_3_70b_instruct_template_arena_hard"
|
| 298 |
+
},
|
| 299 |
+
"score": 0.11387900355871886,
|
| 300 |
+
"score_name": "subsets_mean",
|
| 301 |
+
"num_of_instances": 100
|
| 302 |
+
},
|
| 303 |
+
"entity_extraction": {
|
| 304 |
+
"universal_ner_en_ewt": {
|
| 305 |
+
"num_of_instances": 100,
|
| 306 |
+
"f1_Person": 0.851063829787234,
|
| 307 |
+
"f1_Organization": 0.6984126984126985,
|
| 308 |
+
"f1_Location": 0.7346938775510204,
|
| 309 |
+
"f1_macro": 0.7613901352503176,
|
| 310 |
+
"recall_macro": 0.80175983436853,
|
| 311 |
+
"precision_macro": 0.7273015873015872,
|
| 312 |
+
"in_classes_support": 1.0,
|
| 313 |
+
"f1_micro": 0.7547169811320756,
|
| 314 |
+
"recall_micro": 0.8,
|
| 315 |
+
"precision_micro": 0.7142857142857143,
|
| 316 |
+
"score": 0.7547169811320756,
|
| 317 |
+
"score_name": "f1_micro",
|
| 318 |
+
"score_ci_low": 0.6654206354053237,
|
| 319 |
+
"score_ci_high": 0.8220683069580377,
|
| 320 |
+
"f1_micro_ci_low": 0.6654206354053237,
|
| 321 |
+
"f1_micro_ci_high": 0.8220683069580377
|
| 322 |
+
},
|
| 323 |
+
"score": 0.7547169811320756,
|
| 324 |
+
"score_name": "subsets_mean",
|
| 325 |
+
"num_of_instances": 100
|
| 326 |
+
},
|
| 327 |
+
"knowledge": {
|
| 328 |
+
"mmlu_pro_biology": {
|
| 329 |
+
"accuracy": 0.2857142857142857,
|
| 330 |
+
"accuracy_ci_low": 0.0,
|
| 331 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 332 |
+
"score_name": "accuracy",
|
| 333 |
+
"score": 0.2857142857142857,
|
| 334 |
+
"score_ci_high": 0.7142857142857143,
|
| 335 |
+
"score_ci_low": 0.0,
|
| 336 |
+
"num_of_instances": 7
|
| 337 |
+
},
|
| 338 |
+
"mmlu_pro_business": {
|
| 339 |
+
"accuracy": 0.42857142857142855,
|
| 340 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 341 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 342 |
+
"score_name": "accuracy",
|
| 343 |
+
"score": 0.42857142857142855,
|
| 344 |
+
"score_ci_high": 0.8571428571428571,
|
| 345 |
+
"score_ci_low": 0.14285714285714285,
|
| 346 |
+
"num_of_instances": 7
|
| 347 |
+
},
|
| 348 |
+
"mmlu_pro_chemistry": {
|
| 349 |
+
"accuracy": 0.2857142857142857,
|
| 350 |
+
"accuracy_ci_low": 0.0,
|
| 351 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 352 |
+
"score_name": "accuracy",
|
| 353 |
+
"score": 0.2857142857142857,
|
| 354 |
+
"score_ci_high": 0.7142857142857143,
|
| 355 |
+
"score_ci_low": 0.0,
|
| 356 |
+
"num_of_instances": 7
|
| 357 |
+
},
|
| 358 |
+
"mmlu_pro_computer_science": {
|
| 359 |
+
"accuracy": 0.7142857142857143,
|
| 360 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 361 |
+
"accuracy_ci_high": 1.0,
|
| 362 |
+
"score_name": "accuracy",
|
| 363 |
+
"score": 0.7142857142857143,
|
| 364 |
+
"score_ci_high": 1.0,
|
| 365 |
+
"score_ci_low": 0.2857142857142857,
|
| 366 |
+
"num_of_instances": 7
|
| 367 |
+
},
|
| 368 |
+
"mmlu_pro_economics": {
|
| 369 |
+
"accuracy": 0.5714285714285714,
|
| 370 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 371 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 372 |
+
"score_name": "accuracy",
|
| 373 |
+
"score": 0.5714285714285714,
|
| 374 |
+
"score_ci_high": 0.8571428571428571,
|
| 375 |
+
"score_ci_low": 0.14285714285714285,
|
| 376 |
+
"num_of_instances": 7
|
| 377 |
+
},
|
| 378 |
+
"mmlu_pro_engineering": {
|
| 379 |
+
"accuracy": 0.2857142857142857,
|
| 380 |
+
"accuracy_ci_low": 0.0,
|
| 381 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 382 |
+
"score_name": "accuracy",
|
| 383 |
+
"score": 0.2857142857142857,
|
| 384 |
+
"score_ci_high": 0.7142857142857143,
|
| 385 |
+
"score_ci_low": 0.0,
|
| 386 |
+
"num_of_instances": 7
|
| 387 |
+
},
|
| 388 |
+
"mmlu_pro_health": {
|
| 389 |
+
"accuracy": 0.5714285714285714,
|
| 390 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 391 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 392 |
+
"score_name": "accuracy",
|
| 393 |
+
"score": 0.5714285714285714,
|
| 394 |
+
"score_ci_high": 0.8571428571428571,
|
| 395 |
+
"score_ci_low": 0.14285714285714285,
|
| 396 |
+
"num_of_instances": 7
|
| 397 |
+
},
|
| 398 |
+
"mmlu_pro_history": {
|
| 399 |
+
"accuracy": 0.0,
|
| 400 |
+
"accuracy_ci_low": 0.0,
|
| 401 |
+
"accuracy_ci_high": 0.0,
|
| 402 |
+
"score_name": "accuracy",
|
| 403 |
+
"score": 0.0,
|
| 404 |
+
"score_ci_high": 0.0,
|
| 405 |
+
"score_ci_low": 0.0,
|
| 406 |
+
"num_of_instances": 7
|
| 407 |
+
},
|
| 408 |
+
"mmlu_pro_law": {
|
| 409 |
+
"accuracy": 0.2857142857142857,
|
| 410 |
+
"accuracy_ci_low": 0.0,
|
| 411 |
+
"accuracy_ci_high": 0.7142857142857143,
|
| 412 |
+
"score_name": "accuracy",
|
| 413 |
+
"score": 0.2857142857142857,
|
| 414 |
+
"score_ci_high": 0.7142857142857143,
|
| 415 |
+
"score_ci_low": 0.0,
|
| 416 |
+
"num_of_instances": 7
|
| 417 |
+
},
|
| 418 |
+
"mmlu_pro_math": {
|
| 419 |
+
"accuracy": 0.7142857142857143,
|
| 420 |
+
"accuracy_ci_low": 0.2857142857142857,
|
| 421 |
+
"accuracy_ci_high": 1.0,
|
| 422 |
+
"score_name": "accuracy",
|
| 423 |
+
"score": 0.7142857142857143,
|
| 424 |
+
"score_ci_high": 1.0,
|
| 425 |
+
"score_ci_low": 0.2857142857142857,
|
| 426 |
+
"num_of_instances": 7
|
| 427 |
+
},
|
| 428 |
+
"mmlu_pro_other": {
|
| 429 |
+
"accuracy": 0.42857142857142855,
|
| 430 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 431 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 432 |
+
"score_name": "accuracy",
|
| 433 |
+
"score": 0.42857142857142855,
|
| 434 |
+
"score_ci_high": 0.8571428571428571,
|
| 435 |
+
"score_ci_low": 0.14285714285714285,
|
| 436 |
+
"num_of_instances": 7
|
| 437 |
+
},
|
| 438 |
+
"mmlu_pro_philosophy": {
|
| 439 |
+
"accuracy": 0.14285714285714285,
|
| 440 |
+
"accuracy_ci_low": 0.0,
|
| 441 |
+
"accuracy_ci_high": 0.5714285714285714,
|
| 442 |
+
"score_name": "accuracy",
|
| 443 |
+
"score": 0.14285714285714285,
|
| 444 |
+
"score_ci_high": 0.5714285714285714,
|
| 445 |
+
"score_ci_low": 0.0,
|
| 446 |
+
"num_of_instances": 7
|
| 447 |
+
},
|
| 448 |
+
"mmlu_pro_physics": {
|
| 449 |
+
"accuracy": 0.42857142857142855,
|
| 450 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 451 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 452 |
+
"score_name": "accuracy",
|
| 453 |
+
"score": 0.42857142857142855,
|
| 454 |
+
"score_ci_high": 0.8571428571428571,
|
| 455 |
+
"score_ci_low": 0.14285714285714285,
|
| 456 |
+
"num_of_instances": 7
|
| 457 |
+
},
|
| 458 |
+
"mmlu_pro_psychology": {
|
| 459 |
+
"accuracy": 0.5714285714285714,
|
| 460 |
+
"accuracy_ci_low": 0.14285714285714285,
|
| 461 |
+
"accuracy_ci_high": 0.8571428571428571,
|
| 462 |
+
"score_name": "accuracy",
|
| 463 |
+
"score": 0.5714285714285714,
|
| 464 |
+
"score_ci_high": 0.8571428571428571,
|
| 465 |
+
"score_ci_low": 0.14285714285714285,
|
| 466 |
+
"num_of_instances": 7
|
| 467 |
+
},
|
| 468 |
+
"score": 0.40816326530612246,
|
| 469 |
+
"score_name": "subsets_mean",
|
| 470 |
+
"num_of_instances": 98
|
| 471 |
+
},
|
| 472 |
+
"legal": {
|
| 473 |
+
"legalbench_abercrombie": {
|
| 474 |
+
"f1_macro": 0.25142857142857145,
|
| 475 |
+
"f1_suggestive": 0.2857142857142857,
|
| 476 |
+
"f1_generic": 0.0,
|
| 477 |
+
"f1_descriptive": 0.5714285714285714,
|
| 478 |
+
"f1_fanciful": 0.4,
|
| 479 |
+
"f1_arbitrary": 0.0,
|
| 480 |
+
"f1_macro_ci_low": 0.11666666666666665,
|
| 481 |
+
"f1_macro_ci_high": 0.5085323419170098,
|
| 482 |
+
"score_name": "f1_micro",
|
| 483 |
+
"score": 0.32,
|
| 484 |
+
"score_ci_high": 0.5714285714285714,
|
| 485 |
+
"score_ci_low": 0.09523809523809523,
|
| 486 |
+
"num_of_instances": 20,
|
| 487 |
+
"accuracy": 0.2,
|
| 488 |
+
"accuracy_ci_low": 0.1,
|
| 489 |
+
"accuracy_ci_high": 0.41588290860245253,
|
| 490 |
+
"f1_micro": 0.32,
|
| 491 |
+
"f1_micro_ci_low": 0.09523809523809523,
|
| 492 |
+
"f1_micro_ci_high": 0.5714285714285714
|
| 493 |
+
},
|
| 494 |
+
"legalbench_corporate_lobbying": {
|
| 495 |
+
"f1_macro": 0.5138888888888888,
|
| 496 |
+
"f1_no": 0.5833333333333334,
|
| 497 |
+
"f1_yes": 0.4444444444444444,
|
| 498 |
+
"f1_macro_ci_low": 0.2870981709247007,
|
| 499 |
+
"f1_macro_ci_high": 0.7991422752871308,
|
| 500 |
+
"score_name": "f1_micro",
|
| 501 |
+
"score": 0.5454545454545454,
|
| 502 |
+
"score_ci_high": 0.7428571428571429,
|
| 503 |
+
"score_ci_low": 0.3125,
|
| 504 |
+
"num_of_instances": 20,
|
| 505 |
+
"accuracy": 0.45,
|
| 506 |
+
"accuracy_ci_low": 0.25,
|
| 507 |
+
"accuracy_ci_high": 0.65,
|
| 508 |
+
"f1_micro": 0.5454545454545454,
|
| 509 |
+
"f1_micro_ci_low": 0.3125,
|
| 510 |
+
"f1_micro_ci_high": 0.7428571428571429
|
| 511 |
+
},
|
| 512 |
+
"legalbench_function_of_decision_section": {
|
| 513 |
+
"f1_macro": 0.40816326530612246,
|
| 514 |
+
"f1_conclusion": 0.2857142857142857,
|
| 515 |
+
"f1_issue": 0.25,
|
| 516 |
+
"f1_decree": 0.0,
|
| 517 |
+
"f1_rule": 0.0,
|
| 518 |
+
"f1_analysis": 0.5714285714285714,
|
| 519 |
+
"f1_facts": 0.75,
|
| 520 |
+
"f1_procedural history": 1.0,
|
| 521 |
+
"f1_macro_ci_low": 0.24756971939371533,
|
| 522 |
+
"f1_macro_ci_high": 0.6223765832019531,
|
| 523 |
+
"score_name": "f1_micro",
|
| 524 |
+
"score": 0.42105263157894735,
|
| 525 |
+
"score_ci_high": 0.6153846153846154,
|
| 526 |
+
"score_ci_low": 0.17142857142857143,
|
| 527 |
+
"num_of_instances": 20,
|
| 528 |
+
"accuracy": 0.4,
|
| 529 |
+
"accuracy_ci_low": 0.2,
|
| 530 |
+
"accuracy_ci_high": 0.6,
|
| 531 |
+
"f1_micro": 0.42105263157894735,
|
| 532 |
+
"f1_micro_ci_low": 0.17142857142857143,
|
| 533 |
+
"f1_micro_ci_high": 0.6153846153846154
|
| 534 |
+
},
|
| 535 |
+
"legalbench_international_citizenship_questions": {
|
| 536 |
+
"f1_macro": 0.5248868778280543,
|
| 537 |
+
"f1_yes": 0.46153846153846156,
|
| 538 |
+
"f1_no": 0.5882352941176471,
|
| 539 |
+
"f1_macro_ci_low": 0.2857142857142857,
|
| 540 |
+
"f1_macro_ci_high": 0.7529963905333797,
|
| 541 |
+
"score_name": "f1_micro",
|
| 542 |
+
"score": 0.5333333333333333,
|
| 543 |
+
"score_ci_high": 0.75,
|
| 544 |
+
"score_ci_low": 0.2857142857142857,
|
| 545 |
+
"num_of_instances": 20,
|
| 546 |
+
"accuracy": 0.4,
|
| 547 |
+
"accuracy_ci_low": 0.2,
|
| 548 |
+
"accuracy_ci_high": 0.65,
|
| 549 |
+
"f1_micro": 0.5333333333333333,
|
| 550 |
+
"f1_micro_ci_low": 0.2857142857142857,
|
| 551 |
+
"f1_micro_ci_high": 0.75
|
| 552 |
+
},
|
| 553 |
+
"legalbench_proa": {
|
| 554 |
+
"f1_macro": 1.0,
|
| 555 |
+
"f1_yes": 1.0,
|
| 556 |
+
"f1_no": 1.0,
|
| 557 |
+
"f1_macro_ci_low": 1.0,
|
| 558 |
+
"f1_macro_ci_high": 1.0,
|
| 559 |
+
"score_name": "f1_micro",
|
| 560 |
+
"score": 1.0,
|
| 561 |
+
"score_ci_high": 1.0,
|
| 562 |
+
"score_ci_low": 1.0,
|
| 563 |
+
"num_of_instances": 20,
|
| 564 |
+
"accuracy": 1.0,
|
| 565 |
+
"accuracy_ci_low": 1.0,
|
| 566 |
+
"accuracy_ci_high": 1.0,
|
| 567 |
+
"f1_micro": 1.0,
|
| 568 |
+
"f1_micro_ci_low": 1.0,
|
| 569 |
+
"f1_micro_ci_high": 1.0
|
| 570 |
+
},
|
| 571 |
+
"score": 0.5639681020733652,
|
| 572 |
+
"score_name": "subsets_mean",
|
| 573 |
+
"num_of_instances": 100
|
| 574 |
+
},
|
| 575 |
+
"news_classification": {
|
| 576 |
+
"20_newsgroups_short": {
|
| 577 |
+
"f1_macro": 0.6391478619419796,
|
| 578 |
+
"f1_cars": 0.6666666666666666,
|
| 579 |
+
"f1_windows x": 0.75,
|
| 580 |
+
"f1_computer graphics": 0.6666666666666666,
|
| 581 |
+
"f1_atheism": 0.2857142857142857,
|
| 582 |
+
"f1_religion": 0.0,
|
| 583 |
+
"f1_medicine": 0.8571428571428571,
|
| 584 |
+
"f1_christianity": 0.75,
|
| 585 |
+
"f1_microsoft windows": 0.6,
|
| 586 |
+
"f1_middle east": 0.8,
|
| 587 |
+
"f1_motorcycles": 0.6,
|
| 588 |
+
"f1_for sale": 0.8,
|
| 589 |
+
"f1_mac hardware": 0.8,
|
| 590 |
+
"f1_electronics": 0.4,
|
| 591 |
+
"f1_guns": 0.5454545454545454,
|
| 592 |
+
"f1_politics": 0.5882352941176471,
|
| 593 |
+
"f1_space": 0.6,
|
| 594 |
+
"f1_pc hardware": 0.9230769230769231,
|
| 595 |
+
"f1_cryptography": 0.4,
|
| 596 |
+
"f1_baseball": 1.0,
|
| 597 |
+
"f1_hockey": 0.75,
|
| 598 |
+
"f1_macro_ci_low": 0.5493287854515565,
|
| 599 |
+
"f1_macro_ci_high": 0.7547100417696575,
|
| 600 |
+
"score_name": "f1_micro",
|
| 601 |
+
"score": 0.6736842105263158,
|
| 602 |
+
"score_ci_high": 0.7637409624905424,
|
| 603 |
+
"score_ci_low": 0.5646836389671384,
|
| 604 |
+
"num_of_instances": 100,
|
| 605 |
+
"accuracy": 0.64,
|
| 606 |
+
"accuracy_ci_low": 0.53,
|
| 607 |
+
"accuracy_ci_high": 0.730602617171536,
|
| 608 |
+
"f1_micro": 0.6736842105263158,
|
| 609 |
+
"f1_micro_ci_low": 0.5646836389671384,
|
| 610 |
+
"f1_micro_ci_high": 0.7637409624905424
|
| 611 |
+
},
|
| 612 |
+
"score": 0.6736842105263158,
|
| 613 |
+
"score_name": "subsets_mean",
|
| 614 |
+
"num_of_instances": 100
|
| 615 |
+
},
|
| 616 |
+
"product_help": {
|
| 617 |
+
"cfpb_product_2023": {
|
| 618 |
+
"f1_macro": 0.7947337190327844,
|
| 619 |
+
"f1_debt collection": 0.8571428571428571,
|
| 620 |
+
"f1_checking or savings account": 0.7,
|
| 621 |
+
"f1_credit reporting or credit repair services or other personal consumer reports": 0.8785046728971962,
|
| 622 |
+
"f1_mortgage": 0.6666666666666666,
|
| 623 |
+
"f1_payday loan or title loan or personal loan": 0.8,
|
| 624 |
+
"f1_credit card or prepaid card": 0.9,
|
| 625 |
+
"f1_student loan": 0.8888888888888888,
|
| 626 |
+
"f1_money transfer or virtual currency or money service": 0.6666666666666666,
|
| 627 |
+
"f1_macro_ci_low": 0.6763457146391301,
|
| 628 |
+
"f1_macro_ci_high": 0.886113882457054,
|
| 629 |
+
"score_name": "f1_micro",
|
| 630 |
+
"score": 0.845360824742268,
|
| 631 |
+
"score_ci_high": 0.900523560209424,
|
| 632 |
+
"score_ci_low": 0.7608187004657483,
|
| 633 |
+
"num_of_instances": 100,
|
| 634 |
+
"accuracy": 0.82,
|
| 635 |
+
"accuracy_ci_low": 0.73,
|
| 636 |
+
"accuracy_ci_high": 0.89,
|
| 637 |
+
"f1_micro": 0.845360824742268,
|
| 638 |
+
"f1_micro_ci_low": 0.7608187004657483,
|
| 639 |
+
"f1_micro_ci_high": 0.900523560209424
|
| 640 |
+
},
|
| 641 |
+
"cfpb_product_watsonx": {
|
| 642 |
+
"f1_macro": 0.8501725327812284,
|
| 643 |
+
"f1_mortgages and loans": 0.8333333333333334,
|
| 644 |
+
"f1_credit card": 0.8571428571428571,
|
| 645 |
+
"f1_debt collection": 0.7777777777777778,
|
| 646 |
+
"f1_credit reporting": 0.782608695652174,
|
| 647 |
+
"f1_retail banking": 1.0,
|
| 648 |
+
"f1_macro_ci_low": 0.7375051051121304,
|
| 649 |
+
"f1_macro_ci_high": 0.9324784759448929,
|
| 650 |
+
"score_name": "f1_micro",
|
| 651 |
+
"score": 0.84,
|
| 652 |
+
"score_ci_high": 0.92,
|
| 653 |
+
"score_ci_low": 0.72,
|
| 654 |
+
"num_of_instances": 50,
|
| 655 |
+
"accuracy": 0.84,
|
| 656 |
+
"accuracy_ci_low": 0.72,
|
| 657 |
+
"accuracy_ci_high": 0.92,
|
| 658 |
+
"f1_micro": 0.84,
|
| 659 |
+
"f1_micro_ci_low": 0.72,
|
| 660 |
+
"f1_micro_ci_high": 0.92
|
| 661 |
+
},
|
| 662 |
+
"score": 0.842680412371134,
|
| 663 |
+
"score_name": "subsets_mean",
|
| 664 |
+
"num_of_instances": 150
|
| 665 |
+
},
|
| 666 |
+
"qa_finance": {
|
| 667 |
+
"fin_qa": {
|
| 668 |
+
"num_of_instances": 100,
|
| 669 |
+
"execution_accuracy": 0.18,
|
| 670 |
+
"program_accuracy": 0.22,
|
| 671 |
+
"score": 0.22,
|
| 672 |
+
"score_name": "program_accuracy",
|
| 673 |
+
"execution_accuracy_ci_low": 0.11,
|
| 674 |
+
"execution_accuracy_ci_high": 0.27,
|
| 675 |
+
"program_accuracy_ci_low": 0.14,
|
| 676 |
+
"program_accuracy_ci_high": 0.31,
|
| 677 |
+
"score_ci_low": 0.14,
|
| 678 |
+
"score_ci_high": 0.31
|
| 679 |
+
},
|
| 680 |
+
"score": 0.22,
|
| 681 |
+
"score_name": "subsets_mean",
|
| 682 |
+
"num_of_instances": 100
|
| 683 |
+
},
|
| 684 |
+
"rag_general": {
|
| 685 |
+
"rag_response_generation_clapnq": {
|
| 686 |
+
"precision": 0.4260637500424887,
|
| 687 |
+
"recall": 0.45400425542149314,
|
| 688 |
+
"f1": 0.4031923023399236,
|
| 689 |
+
"precision_ci_low": 0.38955973535176464,
|
| 690 |
+
"precision_ci_high": 0.46184481694256413,
|
| 691 |
+
"recall_ci_low": 0.41391512274499337,
|
| 692 |
+
"recall_ci_high": 0.4911437864233557,
|
| 693 |
+
"f1_ci_low": 0.37554194581461947,
|
| 694 |
+
"f1_ci_high": 0.4323598953031908,
|
| 695 |
+
"score_name": "f1",
|
| 696 |
+
"score": 0.4031923023399236,
|
| 697 |
+
"score_ci_high": 0.4323598953031908,
|
| 698 |
+
"score_ci_low": 0.37554194581461947,
|
| 699 |
+
"num_of_instances": 100,
|
| 700 |
+
"correctness_f1_bert_score.deberta_large_mnli": 0.6653682267665864,
|
| 701 |
+
"correctness_recall_bert_score.deberta_large_mnli": 0.6862254357337951,
|
| 702 |
+
"correctness_precision_bert_score.deberta_large_mnli": 0.6563393643498421,
|
| 703 |
+
"faithfullness_f1_token_overlap": 0.2862937021631226,
|
| 704 |
+
"faithfullness_recall_token_overlap": 0.21242930084003933,
|
| 705 |
+
"faithfullness_precision_token_overlap": 0.5823166581041721,
|
| 706 |
+
"correctness_f1_token_overlap": 0.4031923023399236,
|
| 707 |
+
"correctness_recall_token_overlap": 0.45400425542149314,
|
| 708 |
+
"correctness_precision_token_overlap": 0.4260637500424887
|
| 709 |
+
},
|
| 710 |
+
"score": 0.4031923023399236,
|
| 711 |
+
"score_name": "subsets_mean",
|
| 712 |
+
"num_of_instances": 100
|
| 713 |
+
},
|
| 714 |
+
"reasoning": {
|
| 715 |
+
"hellaswag": {
|
| 716 |
+
"accuracy": 0.58,
|
| 717 |
+
"accuracy_ci_low": 0.47,
|
| 718 |
+
"accuracy_ci_high": 0.6679392176194802,
|
| 719 |
+
"score_name": "accuracy",
|
| 720 |
+
"score": 0.58,
|
| 721 |
+
"score_ci_high": 0.6679392176194802,
|
| 722 |
+
"score_ci_low": 0.47,
|
| 723 |
+
"num_of_instances": 100
|
| 724 |
+
},
|
| 725 |
+
"openbook_qa": {
|
| 726 |
+
"accuracy": 0.93,
|
| 727 |
+
"accuracy_ci_low": 0.86,
|
| 728 |
+
"accuracy_ci_high": 0.97,
|
| 729 |
+
"score_name": "accuracy",
|
| 730 |
+
"score": 0.93,
|
| 731 |
+
"score_ci_high": 0.97,
|
| 732 |
+
"score_ci_low": 0.86,
|
| 733 |
+
"num_of_instances": 100
|
| 734 |
+
},
|
| 735 |
+
"score": 0.755,
|
| 736 |
+
"score_name": "subsets_mean",
|
| 737 |
+
"num_of_instances": 200
|
| 738 |
+
},
|
| 739 |
+
"safety": {
|
| 740 |
+
"attaq_500": {
|
| 741 |
+
"severity_critical": 1.0,
|
| 742 |
+
"severity_high": 4.0,
|
| 743 |
+
"severity_medium": 0.0,
|
| 744 |
+
"severity_low": 95.0,
|
| 745 |
+
"category_harmful_info": 0.8710465241696628,
|
| 746 |
+
"category_deception": 0.8828403099317743,
|
| 747 |
+
"category_explicit_content": 0.7500417944457797,
|
| 748 |
+
"category_violence": 0.8314866851468289,
|
| 749 |
+
"category_discrimination": 0.7108527511831315,
|
| 750 |
+
"category_substance_abuse": 0.898176846239302,
|
| 751 |
+
"category_pii": 0.8761029497340873,
|
| 752 |
+
"safety": 0.8354820214372336,
|
| 753 |
+
"safety_ci_low": 0.8047529357625838,
|
| 754 |
+
"safety_ci_high": 0.8559896869649413,
|
| 755 |
+
"score_name": "safety",
|
| 756 |
+
"score": 0.8354820214372336,
|
| 757 |
+
"score_ci_high": 0.8559896869649413,
|
| 758 |
+
"score_ci_low": 0.8047529357625838,
|
| 759 |
+
"num_of_instances": 100
|
| 760 |
+
},
|
| 761 |
+
"score": 0.8354820214372336,
|
| 762 |
+
"score_name": "subsets_mean",
|
| 763 |
+
"num_of_instances": 100
|
| 764 |
+
},
|
| 765 |
+
"summarization": {
|
| 766 |
+
"billsum_document_filtered_to_6000_chars": {
|
| 767 |
+
"num_of_instances": 100,
|
| 768 |
+
"rouge2": 0.16138724688060274,
|
| 769 |
+
"rougeL": 0.2495894768003618,
|
| 770 |
+
"score": 0.2495894768003618,
|
| 771 |
+
"score_name": "rougeL",
|
| 772 |
+
"rougeLsum": 0.3256656857503464,
|
| 773 |
+
"rouge1": 0.3942786480937058,
|
| 774 |
+
"rouge2_ci_low": 0.14679416476722132,
|
| 775 |
+
"rouge2_ci_high": 0.17444803793356828,
|
| 776 |
+
"rougeL_ci_low": 0.2342488269236526,
|
| 777 |
+
"rougeL_ci_high": 0.26498742032777356,
|
| 778 |
+
"score_ci_low": 0.2342488269236526,
|
| 779 |
+
"score_ci_high": 0.26498742032777356,
|
| 780 |
+
"rougeLsum_ci_low": 0.30302411885582237,
|
| 781 |
+
"rougeLsum_ci_high": 0.34430915502172305,
|
| 782 |
+
"rouge1_ci_low": 0.3696287263777069,
|
| 783 |
+
"rouge1_ci_high": 0.41414117565433084
|
| 784 |
+
},
|
| 785 |
+
"tldr_document_filtered_to_6000_chars": {
|
| 786 |
+
"num_of_instances": 100,
|
| 787 |
+
"rouge2": 0.014373215770455534,
|
| 788 |
+
"rougeL": 0.0764843323714569,
|
| 789 |
+
"score": 0.0764843323714569,
|
| 790 |
+
"score_name": "rougeL",
|
| 791 |
+
"rougeLsum": 0.08525865711564361,
|
| 792 |
+
"rouge1": 0.10811404630359045,
|
| 793 |
+
"rouge2_ci_low": 0.010104689423152697,
|
| 794 |
+
"rouge2_ci_high": 0.019758335317794116,
|
| 795 |
+
"rougeL_ci_low": 0.06567970645776965,
|
| 796 |
+
"rougeL_ci_high": 0.08875406421447059,
|
| 797 |
+
"score_ci_low": 0.06567970645776965,
|
| 798 |
+
"score_ci_high": 0.08875406421447059,
|
| 799 |
+
"rougeLsum_ci_low": 0.07370118292484597,
|
| 800 |
+
"rougeLsum_ci_high": 0.09930080902612025,
|
| 801 |
+
"rouge1_ci_low": 0.09287465904192743,
|
| 802 |
+
"rouge1_ci_high": 0.12519489571711204
|
| 803 |
+
},
|
| 804 |
+
"score": 0.16303690458590936,
|
| 805 |
+
"score_name": "subsets_mean",
|
| 806 |
+
"num_of_instances": 200
|
| 807 |
+
},
|
| 808 |
+
"translation": {
|
| 809 |
+
"mt_flores_101_ara_eng": {
|
| 810 |
+
"num_of_instances": 6,
|
| 811 |
+
"counts": [
|
| 812 |
+
141,
|
| 813 |
+
97,
|
| 814 |
+
70,
|
| 815 |
+
54
|
| 816 |
+
],
|
| 817 |
+
"totals": [
|
| 818 |
+
198,
|
| 819 |
+
192,
|
| 820 |
+
186,
|
| 821 |
+
180
|
| 822 |
+
],
|
| 823 |
+
"precisions": [
|
| 824 |
+
0.7121212121212122,
|
| 825 |
+
0.5052083333333334,
|
| 826 |
+
0.3763440860215054,
|
| 827 |
+
0.3
|
| 828 |
+
],
|
| 829 |
+
"bp": 0.950749126896934,
|
| 830 |
+
"sys_len": 198,
|
| 831 |
+
"ref_len": 208,
|
| 832 |
+
"sacrebleu": 0.42682380184507024,
|
| 833 |
+
"score": 0.42682380184507024,
|
| 834 |
+
"score_name": "sacrebleu",
|
| 835 |
+
"score_ci_low": 0.18249534058920447,
|
| 836 |
+
"score_ci_high": 0.5555695492421439,
|
| 837 |
+
"sacrebleu_ci_low": 0.18249534058920447,
|
| 838 |
+
"sacrebleu_ci_high": 0.5555695492421439
|
| 839 |
+
},
|
| 840 |
+
"mt_flores_101_deu_eng": {
|
| 841 |
+
"num_of_instances": 6,
|
| 842 |
+
"counts": [
|
| 843 |
+
130,
|
| 844 |
+
79,
|
| 845 |
+
47,
|
| 846 |
+
32
|
| 847 |
+
],
|
| 848 |
+
"totals": [
|
| 849 |
+
204,
|
| 850 |
+
198,
|
| 851 |
+
192,
|
| 852 |
+
186
|
| 853 |
+
],
|
| 854 |
+
"precisions": [
|
| 855 |
+
0.6372549019607843,
|
| 856 |
+
0.39898989898989895,
|
| 857 |
+
0.24479166666666669,
|
| 858 |
+
0.17204301075268816
|
| 859 |
+
],
|
| 860 |
+
"bp": 0.9805831403241088,
|
| 861 |
+
"sys_len": 204,
|
| 862 |
+
"ref_len": 208,
|
| 863 |
+
"sacrebleu": 0.3154362559272254,
|
| 864 |
+
"score": 0.3154362559272254,
|
| 865 |
+
"score_name": "sacrebleu",
|
| 866 |
+
"score_ci_low": 0.18558098989202387,
|
| 867 |
+
"score_ci_high": 0.5388353725784034,
|
| 868 |
+
"sacrebleu_ci_low": 0.18558098989202387,
|
| 869 |
+
"sacrebleu_ci_high": 0.5388353725784034
|
| 870 |
+
},
|
| 871 |
+
"mt_flores_101_eng_ara": {
|
| 872 |
+
"num_of_instances": 6,
|
| 873 |
+
"counts": [
|
| 874 |
+
115,
|
| 875 |
+
62,
|
| 876 |
+
39,
|
| 877 |
+
23
|
| 878 |
+
],
|
| 879 |
+
"totals": [
|
| 880 |
+
198,
|
| 881 |
+
192,
|
| 882 |
+
186,
|
| 883 |
+
180
|
| 884 |
+
],
|
| 885 |
+
"precisions": [
|
| 886 |
+
0.5808080808080809,
|
| 887 |
+
0.32291666666666663,
|
| 888 |
+
0.20967741935483872,
|
| 889 |
+
0.1277777777777778
|
| 890 |
+
],
|
| 891 |
+
"bp": 0.9459594689067654,
|
| 892 |
+
"sys_len": 198,
|
| 893 |
+
"ref_len": 209,
|
| 894 |
+
"sacrebleu": 0.25185759673418917,
|
| 895 |
+
"score": 0.25185759673418917,
|
| 896 |
+
"score_name": "sacrebleu",
|
| 897 |
+
"score_ci_low": 0.13927016777509416,
|
| 898 |
+
"score_ci_high": 0.3490260619035246,
|
| 899 |
+
"sacrebleu_ci_low": 0.13927016777509416,
|
| 900 |
+
"sacrebleu_ci_high": 0.3490260619035246
|
| 901 |
+
},
|
| 902 |
+
"mt_flores_101_eng_deu": {
|
| 903 |
+
"num_of_instances": 6,
|
| 904 |
+
"counts": [
|
| 905 |
+
130,
|
| 906 |
+
83,
|
| 907 |
+
61,
|
| 908 |
+
49
|
| 909 |
+
],
|
| 910 |
+
"totals": [
|
| 911 |
+
195,
|
| 912 |
+
190,
|
| 913 |
+
185,
|
| 914 |
+
180
|
| 915 |
+
],
|
| 916 |
+
"precisions": [
|
| 917 |
+
0.6666666666666667,
|
| 918 |
+
0.43684210526315786,
|
| 919 |
+
0.32972972972972975,
|
| 920 |
+
0.2722222222222222
|
| 921 |
+
],
|
| 922 |
+
"bp": 0.8979038320326344,
|
| 923 |
+
"sys_len": 195,
|
| 924 |
+
"ref_len": 216,
|
| 925 |
+
"sacrebleu": 0.36104268767549647,
|
| 926 |
+
"score": 0.36104268767549647,
|
| 927 |
+
"score_name": "sacrebleu",
|
| 928 |
+
"score_ci_low": 0.1966239306512286,
|
| 929 |
+
"score_ci_high": 0.5546980188091108,
|
| 930 |
+
"sacrebleu_ci_low": 0.1966239306512286,
|
| 931 |
+
"sacrebleu_ci_high": 0.5546980188091108
|
| 932 |
+
},
|
| 933 |
+
"mt_flores_101_eng_fra": {
|
| 934 |
+
"num_of_instances": 6,
|
| 935 |
+
"counts": [
|
| 936 |
+
113,
|
| 937 |
+
79,
|
| 938 |
+
54,
|
| 939 |
+
34
|
| 940 |
+
],
|
| 941 |
+
"totals": [
|
| 942 |
+
158,
|
| 943 |
+
154,
|
| 944 |
+
150,
|
| 945 |
+
146
|
| 946 |
+
],
|
| 947 |
+
"precisions": [
|
| 948 |
+
0.7151898734177214,
|
| 949 |
+
0.512987012987013,
|
| 950 |
+
0.36,
|
| 951 |
+
0.2328767123287671
|
| 952 |
+
],
|
| 953 |
+
"bp": 0.6142570611078176,
|
| 954 |
+
"sys_len": 158,
|
| 955 |
+
"ref_len": 235,
|
| 956 |
+
"sacrebleu": 0.25724043173711714,
|
| 957 |
+
"score": 0.25724043173711714,
|
| 958 |
+
"score_name": "sacrebleu",
|
| 959 |
+
"score_ci_low": 0.01141618661583713,
|
| 960 |
+
"score_ci_high": 0.40276606204285015,
|
| 961 |
+
"sacrebleu_ci_low": 0.01141618661583713,
|
| 962 |
+
"sacrebleu_ci_high": 0.40276606204285015
|
| 963 |
+
},
|
| 964 |
+
"mt_flores_101_eng_kor": {
|
| 965 |
+
"num_of_instances": 6,
|
| 966 |
+
"counts": [
|
| 967 |
+
103,
|
| 968 |
+
60,
|
| 969 |
+
36,
|
| 970 |
+
19
|
| 971 |
+
],
|
| 972 |
+
"totals": [
|
| 973 |
+
188,
|
| 974 |
+
184,
|
| 975 |
+
180,
|
| 976 |
+
176
|
| 977 |
+
],
|
| 978 |
+
"precisions": [
|
| 979 |
+
0.5478723404255319,
|
| 980 |
+
0.32608695652173914,
|
| 981 |
+
0.2,
|
| 982 |
+
0.10795454545454546
|
| 983 |
+
],
|
| 984 |
+
"bp": 0.7229117789342253,
|
| 985 |
+
"sys_len": 188,
|
| 986 |
+
"ref_len": 249,
|
| 987 |
+
"sacrebleu": 0.18015911259729336,
|
| 988 |
+
"score": 0.18015911259729336,
|
| 989 |
+
"score_name": "sacrebleu",
|
| 990 |
+
"score_ci_low": 0.00534394043557254,
|
| 991 |
+
"score_ci_high": 0.2674171764554214,
|
| 992 |
+
"sacrebleu_ci_low": 0.00534394043557254,
|
| 993 |
+
"sacrebleu_ci_high": 0.2674171764554214
|
| 994 |
+
},
|
| 995 |
+
"mt_flores_101_eng_por": {
|
| 996 |
+
"num_of_instances": 6,
|
| 997 |
+
"counts": [
|
| 998 |
+
162,
|
| 999 |
+
118,
|
| 1000 |
+
91,
|
| 1001 |
+
72
|
| 1002 |
+
],
|
| 1003 |
+
"totals": [
|
| 1004 |
+
215,
|
| 1005 |
+
209,
|
| 1006 |
+
203,
|
| 1007 |
+
197
|
| 1008 |
+
],
|
| 1009 |
+
"precisions": [
|
| 1010 |
+
0.7534883720930233,
|
| 1011 |
+
0.5645933014354066,
|
| 1012 |
+
0.4482758620689655,
|
| 1013 |
+
0.36548223350253806
|
| 1014 |
+
],
|
| 1015 |
+
"bp": 0.9679661710923415,
|
| 1016 |
+
"sys_len": 215,
|
| 1017 |
+
"ref_len": 222,
|
| 1018 |
+
"sacrebleu": 0.49735461800994313,
|
| 1019 |
+
"score": 0.49735461800994313,
|
| 1020 |
+
"score_name": "sacrebleu",
|
| 1021 |
+
"score_ci_low": 0.3936015193132162,
|
| 1022 |
+
"score_ci_high": 0.5938365039660292,
|
| 1023 |
+
"sacrebleu_ci_low": 0.3936015193132162,
|
| 1024 |
+
"sacrebleu_ci_high": 0.5938365039660292
|
| 1025 |
+
},
|
| 1026 |
+
"mt_flores_101_eng_ron": {
|
| 1027 |
+
"num_of_instances": 6,
|
| 1028 |
+
"counts": [
|
| 1029 |
+
122,
|
| 1030 |
+
83,
|
| 1031 |
+
61,
|
| 1032 |
+
50
|
| 1033 |
+
],
|
| 1034 |
+
"totals": [
|
| 1035 |
+
181,
|
| 1036 |
+
176,
|
| 1037 |
+
171,
|
| 1038 |
+
166
|
| 1039 |
+
],
|
| 1040 |
+
"precisions": [
|
| 1041 |
+
0.6740331491712708,
|
| 1042 |
+
0.47159090909090906,
|
| 1043 |
+
0.3567251461988304,
|
| 1044 |
+
0.30120481927710846
|
| 1045 |
+
],
|
| 1046 |
+
"bp": 0.7628314075724358,
|
| 1047 |
+
"sys_len": 181,
|
| 1048 |
+
"ref_len": 230,
|
| 1049 |
+
"sacrebleu": 0.3279360384622422,
|
| 1050 |
+
"score": 0.3279360384622422,
|
| 1051 |
+
"score_name": "sacrebleu",
|
| 1052 |
+
"score_ci_low": 0.05614266270511962,
|
| 1053 |
+
"score_ci_high": 0.4307120450834441,
|
| 1054 |
+
"sacrebleu_ci_low": 0.05614266270511962,
|
| 1055 |
+
"sacrebleu_ci_high": 0.4307120450834441
|
| 1056 |
+
},
|
| 1057 |
+
"mt_flores_101_eng_spa": {
|
| 1058 |
+
"num_of_instances": 6,
|
| 1059 |
+
"counts": [
|
| 1060 |
+
159,
|
| 1061 |
+
99,
|
| 1062 |
+
67,
|
| 1063 |
+
46
|
| 1064 |
+
],
|
| 1065 |
+
"totals": [
|
| 1066 |
+
233,
|
| 1067 |
+
227,
|
| 1068 |
+
221,
|
| 1069 |
+
215
|
| 1070 |
+
],
|
| 1071 |
+
"precisions": [
|
| 1072 |
+
0.6824034334763948,
|
| 1073 |
+
0.43612334801762115,
|
| 1074 |
+
0.3031674208144796,
|
| 1075 |
+
0.21395348837209302
|
| 1076 |
+
],
|
| 1077 |
+
"bp": 0.957989506197951,
|
| 1078 |
+
"sys_len": 233,
|
| 1079 |
+
"ref_len": 243,
|
| 1080 |
+
"sacrebleu": 0.35708685408102736,
|
| 1081 |
+
"score": 0.35708685408102736,
|
| 1082 |
+
"score_name": "sacrebleu",
|
| 1083 |
+
"score_ci_low": 0.3134751190784969,
|
| 1084 |
+
"score_ci_high": 0.44237915780333503,
|
| 1085 |
+
"sacrebleu_ci_low": 0.3134751190784969,
|
| 1086 |
+
"sacrebleu_ci_high": 0.44237915780333503
|
| 1087 |
+
},
|
| 1088 |
+
"mt_flores_101_fra_eng": {
|
| 1089 |
+
"num_of_instances": 6,
|
| 1090 |
+
"counts": [
|
| 1091 |
+
144,
|
| 1092 |
+
97,
|
| 1093 |
+
64,
|
| 1094 |
+
44
|
| 1095 |
+
],
|
| 1096 |
+
"totals": [
|
| 1097 |
+
219,
|
| 1098 |
+
213,
|
| 1099 |
+
207,
|
| 1100 |
+
201
|
| 1101 |
+
],
|
| 1102 |
+
"precisions": [
|
| 1103 |
+
0.6575342465753425,
|
| 1104 |
+
0.4553990610328638,
|
| 1105 |
+
0.30917874396135264,
|
| 1106 |
+
0.21890547263681592
|
| 1107 |
+
],
|
| 1108 |
+
"bp": 1.0,
|
| 1109 |
+
"sys_len": 219,
|
| 1110 |
+
"ref_len": 208,
|
| 1111 |
+
"sacrebleu": 0.3773064147727779,
|
| 1112 |
+
"score": 0.3773064147727779,
|
| 1113 |
+
"score_name": "sacrebleu",
|
| 1114 |
+
"score_ci_low": 0.19660284291022648,
|
| 1115 |
+
"score_ci_high": 0.5367507535790375,
|
| 1116 |
+
"sacrebleu_ci_low": 0.19660284291022648,
|
| 1117 |
+
"sacrebleu_ci_high": 0.5367507535790375
|
| 1118 |
+
},
|
| 1119 |
+
"mt_flores_101_jpn_eng": {
|
| 1120 |
+
"num_of_instances": 6,
|
| 1121 |
+
"counts": [
|
| 1122 |
+
131,
|
| 1123 |
+
64,
|
| 1124 |
+
36,
|
| 1125 |
+
23
|
| 1126 |
+
],
|
| 1127 |
+
"totals": [
|
| 1128 |
+
210,
|
| 1129 |
+
204,
|
| 1130 |
+
198,
|
| 1131 |
+
192
|
| 1132 |
+
],
|
| 1133 |
+
"precisions": [
|
| 1134 |
+
0.6238095238095238,
|
| 1135 |
+
0.3137254901960784,
|
| 1136 |
+
0.18181818181818182,
|
| 1137 |
+
0.11979166666666666
|
| 1138 |
+
],
|
| 1139 |
+
"bp": 1.0,
|
| 1140 |
+
"sys_len": 210,
|
| 1141 |
+
"ref_len": 208,
|
| 1142 |
+
"sacrebleu": 0.2555150181574835,
|
| 1143 |
+
"score": 0.2555150181574835,
|
| 1144 |
+
"score_name": "sacrebleu",
|
| 1145 |
+
"score_ci_low": 0.10158659890719782,
|
| 1146 |
+
"score_ci_high": 0.34125387925677286,
|
| 1147 |
+
"sacrebleu_ci_low": 0.10158659890719782,
|
| 1148 |
+
"sacrebleu_ci_high": 0.34125387925677286
|
| 1149 |
+
},
|
| 1150 |
+
"mt_flores_101_kor_eng": {
|
| 1151 |
+
"num_of_instances": 6,
|
| 1152 |
+
"counts": [
|
| 1153 |
+
127,
|
| 1154 |
+
73,
|
| 1155 |
+
45,
|
| 1156 |
+
30
|
| 1157 |
+
],
|
| 1158 |
+
"totals": [
|
| 1159 |
+
201,
|
| 1160 |
+
195,
|
| 1161 |
+
189,
|
| 1162 |
+
183
|
| 1163 |
+
],
|
| 1164 |
+
"precisions": [
|
| 1165 |
+
0.6318407960199005,
|
| 1166 |
+
0.37435897435897436,
|
| 1167 |
+
0.2380952380952381,
|
| 1168 |
+
0.16393442622950818
|
| 1169 |
+
],
|
| 1170 |
+
"bp": 0.9657735711441044,
|
| 1171 |
+
"sys_len": 201,
|
| 1172 |
+
"ref_len": 208,
|
| 1173 |
+
"sacrebleu": 0.29936740534149714,
|
| 1174 |
+
"score": 0.29936740534149714,
|
| 1175 |
+
"score_name": "sacrebleu",
|
| 1176 |
+
"score_ci_low": 0.17940891931864458,
|
| 1177 |
+
"score_ci_high": 0.4579443376656351,
|
| 1178 |
+
"sacrebleu_ci_low": 0.17940891931864458,
|
| 1179 |
+
"sacrebleu_ci_high": 0.4579443376656351
|
| 1180 |
+
},
|
| 1181 |
+
"mt_flores_101_por_eng": {
|
| 1182 |
+
"num_of_instances": 6,
|
| 1183 |
+
"counts": [
|
| 1184 |
+
150,
|
| 1185 |
+
105,
|
| 1186 |
+
74,
|
| 1187 |
+
55
|
| 1188 |
+
],
|
| 1189 |
+
"totals": [
|
| 1190 |
+
201,
|
| 1191 |
+
195,
|
| 1192 |
+
189,
|
| 1193 |
+
183
|
| 1194 |
+
],
|
| 1195 |
+
"precisions": [
|
| 1196 |
+
0.746268656716418,
|
| 1197 |
+
0.5384615384615384,
|
| 1198 |
+
0.3915343915343915,
|
| 1199 |
+
0.3005464480874317
|
| 1200 |
+
],
|
| 1201 |
+
"bp": 0.9657735711441044,
|
| 1202 |
+
"sys_len": 201,
|
| 1203 |
+
"ref_len": 208,
|
| 1204 |
+
"sacrebleu": 0.4503582964904346,
|
| 1205 |
+
"score": 0.4503582964904346,
|
| 1206 |
+
"score_name": "sacrebleu",
|
| 1207 |
+
"score_ci_low": 0.23158737188314713,
|
| 1208 |
+
"score_ci_high": 0.598006434562735,
|
| 1209 |
+
"sacrebleu_ci_low": 0.23158737188314713,
|
| 1210 |
+
"sacrebleu_ci_high": 0.598006434562735
|
| 1211 |
+
},
|
| 1212 |
+
"mt_flores_101_ron_eng": {
|
| 1213 |
+
"num_of_instances": 6,
|
| 1214 |
+
"counts": [
|
| 1215 |
+
152,
|
| 1216 |
+
103,
|
| 1217 |
+
71,
|
| 1218 |
+
50
|
| 1219 |
+
],
|
| 1220 |
+
"totals": [
|
| 1221 |
+
216,
|
| 1222 |
+
210,
|
| 1223 |
+
204,
|
| 1224 |
+
198
|
| 1225 |
+
],
|
| 1226 |
+
"precisions": [
|
| 1227 |
+
0.7037037037037037,
|
| 1228 |
+
0.4904761904761905,
|
| 1229 |
+
0.3480392156862745,
|
| 1230 |
+
0.25252525252525254
|
| 1231 |
+
],
|
| 1232 |
+
"bp": 1.0,
|
| 1233 |
+
"sys_len": 216,
|
| 1234 |
+
"ref_len": 208,
|
| 1235 |
+
"sacrebleu": 0.4173353670626537,
|
| 1236 |
+
"score": 0.4173353670626537,
|
| 1237 |
+
"score_name": "sacrebleu",
|
| 1238 |
+
"score_ci_low": 0.27899206861096376,
|
| 1239 |
+
"score_ci_high": 0.58440086756332,
|
| 1240 |
+
"sacrebleu_ci_low": 0.27899206861096376,
|
| 1241 |
+
"sacrebleu_ci_high": 0.58440086756332
|
| 1242 |
+
},
|
| 1243 |
+
"mt_flores_101_spa_eng": {
|
| 1244 |
+
"num_of_instances": 6,
|
| 1245 |
+
"counts": [
|
| 1246 |
+
97,
|
| 1247 |
+
59,
|
| 1248 |
+
39,
|
| 1249 |
+
26
|
| 1250 |
+
],
|
| 1251 |
+
"totals": [
|
| 1252 |
+
164,
|
| 1253 |
+
159,
|
| 1254 |
+
154,
|
| 1255 |
+
149
|
| 1256 |
+
],
|
| 1257 |
+
"precisions": [
|
| 1258 |
+
0.5914634146341463,
|
| 1259 |
+
0.37106918238993714,
|
| 1260 |
+
0.2532467532467533,
|
| 1261 |
+
0.174496644295302
|
| 1262 |
+
],
|
| 1263 |
+
"bp": 0.764683938413801,
|
| 1264 |
+
"sys_len": 164,
|
| 1265 |
+
"ref_len": 208,
|
| 1266 |
+
"sacrebleu": 0.2399719071849068,
|
| 1267 |
+
"score": 0.2399719071849068,
|
| 1268 |
+
"score_name": "sacrebleu",
|
| 1269 |
+
"score_ci_low": 0.05189527961371486,
|
| 1270 |
+
"score_ci_high": 0.411377459488148,
|
| 1271 |
+
"sacrebleu_ci_low": 0.05189527961371486,
|
| 1272 |
+
"sacrebleu_ci_high": 0.411377459488148
|
| 1273 |
+
},
|
| 1274 |
+
"score": 0.3343194537386239,
|
| 1275 |
+
"score_name": "subsets_mean",
|
| 1276 |
+
"num_of_instances": 90
|
| 1277 |
+
},
|
| 1278 |
+
"score": 0.5398167389664902,
|
| 1279 |
+
"score_name": "subsets_mean",
|
| 1280 |
+
"num_of_instances": 1537
|
| 1281 |
+
}
|
| 1282 |
+
}
|