Upload vocagno_config.json with huggingface_hub
Browse files- vocagno_config.json +27 -0
vocagno_config.json
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"train_file": "data/processed/dolly_train.jsonl,data/processed/gsm8k_train.jsonl,data/processed/dolphin_train.jsonl,data/processed/openhermes_train.jsonl,data/processed/slimorca_train.jsonl,data/processed/metamath_train.jsonl,data/processed/oasst_train.jsonl",
|
| 3 |
+
"valid_file": "data/processed/dolly_train.jsonl",
|
| 4 |
+
"max_seq_length": 1024,
|
| 5 |
+
"prompt_max_length": 512,
|
| 6 |
+
"teacher_name_or_path": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
|
| 7 |
+
"teacher_tokenizer": null,
|
| 8 |
+
"student_checkpoint": "./checkpoints/minillm_enhanced_4b_nothink/final_model/",
|
| 9 |
+
"student_tokenizer": "Qwen/Qwen3-0.6B",
|
| 10 |
+
"output_dir": "checkpoints/vocagno_deepseek",
|
| 11 |
+
"num_steps": 100000,
|
| 12 |
+
"batch_size": 2,
|
| 13 |
+
"gradient_accumulation_steps": 8,
|
| 14 |
+
"learning_rate": 2e-05,
|
| 15 |
+
"weight_decay": 0.01,
|
| 16 |
+
"warmup_ratio": 0.1,
|
| 17 |
+
"max_grad_norm": 1.0,
|
| 18 |
+
"aggregation": "mean",
|
| 19 |
+
"top_k_ratio": 0.4,
|
| 20 |
+
"kd_weight": 1.0,
|
| 21 |
+
"sft_weight": 0.3,
|
| 22 |
+
"full_sequence_mode": true,
|
| 23 |
+
"bf16": true,
|
| 24 |
+
"logging_steps": 25,
|
| 25 |
+
"save_steps": 500,
|
| 26 |
+
"seed": 42
|
| 27 |
+
}
|