27 lines
1009 B
JSON
27 lines
1009 B
JSON
|
{
|
||
|
|
"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",
|
||
|
|
"valid_file": "data/processed/dolly_train.jsonl",
|
||
|
|
"max_seq_length": 1024,
|
||
|
|
"prompt_max_length": 512,
|
||
|
|
"teacher_name_or_path": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
|
||
|
|
"teacher_tokenizer": null,
|
||
|
|
"student_checkpoint": "./checkpoints/minillm_enhanced_4b_nothink/final_model/",
|
||
|
|
"student_tokenizer": "Qwen/Qwen3-0.6B",
|
||
|
|
"output_dir": "checkpoints/vocagno_deepseek",
|
||
|
|
"num_steps": 100000,
|
||
|
|
"batch_size": 2,
|
||
|
|
"gradient_accumulation_steps": 8,
|
||
|
|
"learning_rate": 2e-05,
|
||
|
|
"weight_decay": 0.01,
|
||
|
|
"warmup_ratio": 0.1,
|
||
|
|
"max_grad_norm": 1.0,
|
||
|
|
"aggregation": "mean",
|
||
|
|
"top_k_ratio": 0.4,
|
||
|
|
"kd_weight": 1.0,
|
||
|
|
"sft_weight": 0.3,
|
||
|
|
"full_sequence_mode": true,
|
||
|
|
"bf16": true,
|
||
|
|
"logging_steps": 25,
|
||
|
|
"save_steps": 500,
|
||
|
|
"seed": 42
|
||
|
|
}
|