{
  "run_dirs": [
    "data_pipelines/model_understanding/runs/qwen3_8b_50k_train_structured_filtered"
  ],
  "model_name": "Qwen/Qwen3-8B",
  "prepared_dataset_path": "text_sft_training_data/mu_qwen3_8b_50k_structured_answer_filtered.pt",
  "save_dir": "checkpoints_text_sft/mu_qwen3_8b_50k_structured_answer_e3_kl0_filtered",
  "wandb_project": "model_understanding_sft",
  "wandb_run_name": "mu_qwen3_8b_50k_structured_answer_e3_kl0_filtered",
  "global_train_batch_size": 8,
  "num_epochs": 3,
  "lr": 5e-05,
  "schema_version": 1,
  "min_interest_score": 3,
  "min_verification_score": 7,
  "completion_selection_mode": "first",
  "synthetic_data_paths": null,
  "max_train_examples": null,
  "max_seq_len": 4096,
  "max_steps": null,
  "gradient_accumulation_steps": 1,
  "warmup_ratio": 0.05,
  "max_grad_norm": 1.0,
  "gradient_checkpointing": true,
  "window_mult": 20,
  "use_lora": true,
  "load_lora_path": null,
  "lora_r": 64,
  "lora_alpha": 128,
  "lora_dropout": 0.05,
  "lora_target_modules": "all-linear",
  "load_in_8bit": false,
  "save_steps": 9999999,
  "log_steps": 1,
  "debug_num_examples_to_dump": 2,
  "seed": 42,
  "kl_loss_weight": 0.0,
  "kl_every_n_steps": 100,
  "kl_batch_size": 2,
  "kl_data_path": null
}