benchmarks/verify_training.json

{
  "checks": {
    "overfit_ok": true,
    "reference_loss_ok": true,
    "refine_numerics_ok": true
  },
  "config": {
    "batch_size": 32,
    "d_model": 128,
    "device": "cuda:0",
    "n_tail": 4,
    "overfit_steps": 120,
    "prefix_len": 4,
    "refine_iters": 2,
    "vocab_size": 64
  },
  "overfit": {
    "final_loss": 0.0031353291124105453,
    "initial_loss": 4.357937812805176,
    "loss_ratio": 0.0007194524674486704,
    "tail_token_accuracy": 1.0
  },
  "passed": true,
  "real_batch_sanity": null,
  "reference_loss": {
    "abs_diff": 0.0,
    "manual_loss": 4.265489101409912,
    "training_loss": 4.265489101409912
  },
  "refine_stability": {
    "changed_token_fraction_iter2_vs_iter1": 0.1875,
    "finite_logits": true,
    "max_logit_delta": 0.19519612193107605,
    "token_bounds_ok": true
  }
}