{
  "run_slug": "opsd-llama31-8b-instruct-origin-gen1024-step200-jsdclip006-20260515",
  "method": "OPSD origin fixed-teacher full-vocabulary JSD with per-token clipping",
  "base_model": "meta-llama/Llama-3.1-8B-Instruct",
  "model_family": "llama",
  "dataset": "siyanzhao/Openthoughts_math_30k_opsd",
  "max_steps": 200,
  "max_completion_length": 1024,
  "max_length": 20000,
  "student_thinking": false,
  "teacher_thinking": true,
  "fixed_teacher": true,
  "loss": "full-vocabulary forward KL/JSD beta=0",
  "beta": 0.0,
  "lmbda": 1.0,
  "jsd_token_clip": 0.06,
  "temperature": 1.1,
  "top_p": 0.95,
  "top_k": 20,
  "per_device_train_batch_size": 1,
  "gradient_accumulation_steps": 2,
  "effective_train_batch_size": 8,
  "learning_rate": 5e-06,
  "max_grad_norm": 0.1,
  "vllm_mode": "colocate",
  "vllm_gpu_memory_utilization": 0.35,
  "vllm_tensor_parallel_size": 1,
  "lora_r": 64,
  "lora_alpha": 128,
  "lora_target_modules": [
    "q_proj",
    "k_proj",
    "v_proj",
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj"
  ],
  "num_gpus": 4,
  "uploaded_artifacts": "LoRA adapter checkpoints, tokenizer files, trainer state, generations, logs, README, metadata"
}