Upload configs.yaml with huggingface_hub
Browse files- configs.yaml +40 -0
configs.yaml
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
assistant_tag: gpt
|
| 2 |
+
bf16: 'True'
|
| 3 |
+
content_tag: value
|
| 4 |
+
cutoff_len: '16384'
|
| 5 |
+
dataset: mlfoundations-dev/wikipedia_seed_science
|
| 6 |
+
dataset_dir: ONLINE
|
| 7 |
+
ddp_timeout: '180000000'
|
| 8 |
+
deepspeed: /opt/ml/code/zero3.json
|
| 9 |
+
do_train: 'True'
|
| 10 |
+
enable_liger_kernel: 'True'
|
| 11 |
+
finetuning_type: full
|
| 12 |
+
formatting: sharegpt
|
| 13 |
+
global_batch_size: '96'
|
| 14 |
+
gradient_accumulation_steps: '3'
|
| 15 |
+
hub_model_id: mlfoundations-dev/ds_no_offload_liger_packing_torchcompile
|
| 16 |
+
learning_rate: 1e-05
|
| 17 |
+
logging_steps: '1'
|
| 18 |
+
lr_scheduler_type: cosine
|
| 19 |
+
max_samples: '1000000'
|
| 20 |
+
messages: conversations
|
| 21 |
+
model_name_or_path: Qwen/Qwen2.5-7B-Instruct
|
| 22 |
+
neat_packing: 'True'
|
| 23 |
+
num_train_epochs: '3.0'
|
| 24 |
+
output_dir: /opt/ml/model
|
| 25 |
+
overwrite_cache: 'True'
|
| 26 |
+
packing: 'True'
|
| 27 |
+
per_device_train_batch_size: '1'
|
| 28 |
+
plot_loss: 'True'
|
| 29 |
+
preprocessing_num_workers: '16'
|
| 30 |
+
push_to_db: 'True'
|
| 31 |
+
push_to_hub: 'True'
|
| 32 |
+
report_to: wandb
|
| 33 |
+
role_tag: from
|
| 34 |
+
run_name: ds_no_offload_liger_packing_torchcompile
|
| 35 |
+
save_strategy: epoch
|
| 36 |
+
stage: sft
|
| 37 |
+
template: qwen25
|
| 38 |
+
torch_compile: 'True'
|
| 39 |
+
user_tag: human
|
| 40 |
+
warmup_ratio: '0.1'
|