File size: 4,085 Bytes
4a78d34 bbde2b0 4a78d34 bbde2b0 a2189ab 36244aa bbde2b0 7c6bd6c bbde2b0 cd53742 bbde2b0 7c6bd6c 4a78d34 dea22be a2f2df3 4a78d34 18638a9 4a78d34 18638a9 d7d56ae bbde2b0 ca19cea d7d56ae cdca101 36244aa 4a78d34 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 |
import json
import os
import pandas as pd
from src.about import Tasks
from src.display.formatting import has_no_nan_values, make_clickable_model
from src.display.utils import AutoEvalColumn, EvalQueueColumn
from src.leaderboard.read_evals import get_raw_eval_results
TASK_NAME_INVERSE_MAP = dict()
for task in Tasks:
TASK_NAME_INVERSE_MAP[task.value.col_name] = {
"name": task.value.benchmark,
"type": task.value.type,
"source": task.value.source,
}
def get_inspect_log_url(model_name: str, benchmark_name: str) -> str:
"""Returns the URL to the log file for a given model and benchmark"""
with open("./inspect_log_file_names.json", "r") as f:
inspect_log_files = json.load(f)
log_file_name = inspect_log_files[model_name].get(benchmark_name, None)
# replace .json with .eval
log_file_name = log_file_name.replace(".json", ".eval")
if log_file_name is None:
return ""
else:
return f"https://storage.googleapis.com/inspect-evals/{model_name}/index.html?log_file=logs/logs/{log_file_name}"
def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame:
"""Creates a dataframe from all the individual experiment results"""
raw_data = get_raw_eval_results(results_path, requests_path)
all_data_json = [v.to_dict() for v in raw_data]
df = pd.DataFrame.from_records(all_data_json)
# df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
df = df[cols].round(decimals=2)
# TMP: Debug
print(df.shape())
# filter out if any of the benchmarks have not been produced
df = df[has_no_nan_values(df, benchmark_cols)]
# TMP: Debug
print(df.shape())
# make values clickable and link to log files
for col in benchmark_cols:
df[col] = df[[AutoEvalColumn.model.name, col]].apply(lambda x: f"[{x[col]}]({get_inspect_log_url(model_name=x[AutoEvalColumn.model.name].split('>')[1].split('<')[0], benchmark_name=TASK_NAME_INVERSE_MAP[col]['name'])})", axis=1)
# # make task names clickable and link to inspect-evals repository - this creates issues later
# df = df.rename(columns={col: f"[{col}]({TASK_NAME_INVERSE_MAP[col]['source']})" for col in benchmark_cols})
return df
def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
"""Creates the different dataframes for the evaluation queues requestes"""
entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")]
all_evals = []
for entry in entries:
if ".json" in entry:
file_path = os.path.join(save_path, entry)
with open(file_path) as fp:
data = json.load(fp)
data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
data[EvalQueueColumn.revision.name] = data.get("revision", "main")
all_evals.append(data)
elif ".md" not in entry:
# this is a folder
sub_entries = [e for e in os.listdir(f"{save_path}/{entry}") if os.path.isfile(e) and not e.startswith(".")]
for sub_entry in sub_entries:
file_path = os.path.join(save_path, entry, sub_entry)
with open(file_path) as fp:
data = json.load(fp)
data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
data[EvalQueueColumn.revision.name] = data.get("revision", "main")
all_evals.append(data)
pending_list = [e for e in all_evals if e["status"] in ["PENDING", "RERUN"]]
running_list = [e for e in all_evals if e["status"] == "RUNNING"]
finished_list = [e for e in all_evals if e["status"].startswith("FINISHED") or e["status"] == "PENDING_NEW_EVAL"]
df_pending = pd.DataFrame.from_records(pending_list, columns=cols)
df_running = pd.DataFrame.from_records(running_list, columns=cols)
df_finished = pd.DataFrame.from_records(finished_list, columns=cols)
return df_finished[cols], df_running[cols], df_pending[cols]
|