train_state.json

{
  "global_step": 900,
  "data_position": 14546,
  "data_state": {
    "variant": "streaming_token_stream",
    "consumed": 14546,
    "dataset": "karpathy/climbmix-400b-shuffle",
    "dataset_split": "train",
    "max_seq_len": 1200,
    "min_chars": 200,
    "teacher_model": "Qwen/Qwen3.5-35B-A3B"
  },
  "best_beating_delta": -Infinity,
  "beat_streak": 0,
  "gate_delta_ema": 0.02357170544715706,
  "gate_streak": 0,
  "best_gate_score": -Infinity,
  "best_gate_step": 0
}