107 lines
2.7 KiB
JSON
107 lines
2.7 KiB
JSON
|
{
|
||
|
|
"dataset_ids": [
|
||
|
|
"0xZee/dataset-CoT-Advanced-Calculus-268"
|
||
|
|
],
|
||
|
|
"dataset_split": "train",
|
||
|
|
"eval_fraction": 0.1,
|
||
|
|
"seed": 42,
|
||
|
|
"max_examples": null,
|
||
|
|
"question_fields": [
|
||
|
|
"question",
|
||
|
|
"instruction",
|
||
|
|
"problem",
|
||
|
|
"prompt",
|
||
|
|
"input"
|
||
|
|
],
|
||
|
|
"reasoning_fields": [
|
||
|
|
"CoT",
|
||
|
|
"cot",
|
||
|
|
"chain_of_thought",
|
||
|
|
"reasoning",
|
||
|
|
"analysis",
|
||
|
|
"rationale",
|
||
|
|
"solution"
|
||
|
|
],
|
||
|
|
"answer_fields": [
|
||
|
|
"response",
|
||
|
|
"final_answer",
|
||
|
|
"answer",
|
||
|
|
"output",
|
||
|
|
"completion"
|
||
|
|
],
|
||
|
|
"save_data_audit": true,
|
||
|
|
"data_audit_path": "./DualMind100/data_audit.jsonl",
|
||
|
|
"student_id": "Qwen/Qwen3-1.7B",
|
||
|
|
"teacher_id": "Qwen/Qwen3-8B",
|
||
|
|
"student_revision": null,
|
||
|
|
"teacher_revision": null,
|
||
|
|
"save_dir": "./DualMind100",
|
||
|
|
"resume_from": null,
|
||
|
|
"use_bf16": true,
|
||
|
|
"keep_student_fp32": true,
|
||
|
|
"gradient_checkpointing": true,
|
||
|
|
"gradient_checkpointing_use_reentrant": false,
|
||
|
|
"attention_implementation": "sdpa",
|
||
|
|
"compile_student": false,
|
||
|
|
"teacher_load_in_4bit": false,
|
||
|
|
"teacher_device_map": "auto",
|
||
|
|
"teacher_offload_folder": "./teacher_offload",
|
||
|
|
"teacher_max_memory_gpu": null,
|
||
|
|
"teacher_max_memory_cpu": null,
|
||
|
|
"use_8bit_optimizer": false,
|
||
|
|
"use_paged_optimizer": true,
|
||
|
|
"dual_mind_format": true,
|
||
|
|
"use_chat_template": true,
|
||
|
|
"system_prompt": "Solve the problem rigorously. Preserve assumptions, derivation, verification, and final result.",
|
||
|
|
"preserve_full_answer": true,
|
||
|
|
"window_size": 2048,
|
||
|
|
"min_overlap": 128,
|
||
|
|
"max_overlap": 384,
|
||
|
|
"teacher_top_k": 64,
|
||
|
|
"teacher_cache_dir": "./tkd_cache_v4",
|
||
|
|
"teacher_context_tokens": 512,
|
||
|
|
"vocab_logsumexp_chunk": 16384,
|
||
|
|
"cache_log_every": 20,
|
||
|
|
"min_expected_topk_mass": 0.85,
|
||
|
|
"batch_size": 2,
|
||
|
|
"grad_accum_steps": 4,
|
||
|
|
"epochs": 3,
|
||
|
|
"lr": 3e-06,
|
||
|
|
"min_lr": 3e-07,
|
||
|
|
"warmup_ratio": 0.05,
|
||
|
|
"weight_decay": 0.01,
|
||
|
|
"adam_beta1": 0.9,
|
||
|
|
"adam_beta2": 0.95,
|
||
|
|
"adam_eps": 1e-08,
|
||
|
|
"grad_clip": 1.0,
|
||
|
|
"temperature": 2.0,
|
||
|
|
"kd_target_alpha": 0.35,
|
||
|
|
"kd_start_frac": 0.1,
|
||
|
|
"kd_end_frac": 0.4,
|
||
|
|
"eval_kd_alpha": null,
|
||
|
|
"jump_threshold_sigma": 3.5,
|
||
|
|
"jump_threshold_quantile": 0.995,
|
||
|
|
"support_turnover_weight": 0.15,
|
||
|
|
"jump_amplifier": 1.33,
|
||
|
|
"topology_energy_window": 32,
|
||
|
|
"topology_pair_chunk": 512,
|
||
|
|
"dg_window": 3,
|
||
|
|
"gap_energy_alpha": 0.0,
|
||
|
|
"proof_weight_start": 2.0,
|
||
|
|
"proof_weight_end": 1.1,
|
||
|
|
"examine_extra_multiplier": 1.15,
|
||
|
|
"response_weight": 1.0,
|
||
|
|
"curriculum_enabled": true,
|
||
|
|
"curriculum_phases": 4,
|
||
|
|
"easy_first_fraction": 0.3,
|
||
|
|
"num_workers": 2,
|
||
|
|
"prefetch_factor": 2,
|
||
|
|
"log_every": 10,
|
||
|
|
"eval_every": 100,
|
||
|
|
"checkpoint_every_opt_steps": 250,
|
||
|
|
"save_every_epoch": true,
|
||
|
|
"save_best": true,
|
||
|
|
"best_metric": "ce",
|
||
|
|
"run_internal_checks": true,
|
||
|
|
"inference_max_new_tokens": 768
|
||
|
|
}
|