30 lines
899 B
JSON
30 lines
899 B
JSON
{
|
|
"stage": "SFT",
|
|
"base_model": "Qwen/Qwen3-4B-Base",
|
|
"source_checkpoint_name": "qwen3-4b-recursive-sft-v3.2-balanced",
|
|
"component_examples": 14126,
|
|
"decomposition_examples": 4707,
|
|
"direct_examples": 4707,
|
|
"aggregation_examples": 4712,
|
|
"unique_source_problems": 5241,
|
|
"configured_epochs": 6,
|
|
"optimizer_updates": 660,
|
|
"final_epoch": 5.951302378255946,
|
|
"global_component_batch_size": 128,
|
|
"training_gpu_count": 8,
|
|
"per_device_batch_size": 1,
|
|
"gradient_accumulation_steps": 16,
|
|
"max_sequence_length": 2048,
|
|
"learning_rate": 1e-05,
|
|
"weight_decay": 0.01,
|
|
"warmup_ratio": 0.1,
|
|
"scheduler": "cosine",
|
|
"seed": 1,
|
|
"precision": "bf16",
|
|
"gradient_checkpointing": true,
|
|
"loss": "output_tokens_only",
|
|
"prompt_representation": "raw_recursive_strings_not_chat_template",
|
|
"protocol": "original_tags_not_json",
|
|
"known_unrepaired_terminal_answer_targets": 387
|
|
}
|