Upload configs.yaml with huggingface_hub
Browse files- configs.yaml +34 -0
configs.yaml
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
bf16: true
|
| 2 |
+
cutoff_len: 16384
|
| 3 |
+
dataset: mlfoundations-dev/sky_t1_2-5k_rewrite_r1_distill_llama70b
|
| 4 |
+
dataset_dir: ONLINE
|
| 5 |
+
ddp_timeout: 180000000
|
| 6 |
+
deepspeed: dcft/train/zero3.json
|
| 7 |
+
do_train: true
|
| 8 |
+
eval_strategy: 'no'
|
| 9 |
+
finetuning_type: full
|
| 10 |
+
formatting: sharegpt
|
| 11 |
+
global_batch_size: 96
|
| 12 |
+
gradient_accumulation_steps: 3
|
| 13 |
+
hub_model_id: mlfoundations-dev/qwen2-5_sky_t1_2-5k_rewrite_r1_distill_llama70b
|
| 14 |
+
include_hp: dcft/train/hp_settings/reasoning.yaml
|
| 15 |
+
learning_rate: 1.0e-05
|
| 16 |
+
logging_steps: 1
|
| 17 |
+
lr_scheduler_type: cosine
|
| 18 |
+
max_samples: 1000000
|
| 19 |
+
messages: conversations
|
| 20 |
+
model_name_or_path: Qwen/Qwen2.5-7B-Instruct
|
| 21 |
+
num_train_epochs: 3.0
|
| 22 |
+
output_dir: /tmp/dcft_checkpoints/train/checkpoints/qwen2-5_sky_t1_2-5k_rewrite_r1_distill_llama70b
|
| 23 |
+
overwrite_cache: true
|
| 24 |
+
per_device_train_batch_size: 1
|
| 25 |
+
plot_loss: true
|
| 26 |
+
preprocessing_num_workers: 16
|
| 27 |
+
push_to_db: true
|
| 28 |
+
push_to_hub: true
|
| 29 |
+
report_to: wandb
|
| 30 |
+
run_name: qwen2-5_sky_t1_2-5k_rewrite_r1_distill_llama70b
|
| 31 |
+
save_strategy: epoch
|
| 32 |
+
stage: sft
|
| 33 |
+
template: qwen25
|
| 34 |
+
warmup_ratio: 0.1
|