declref_5pct_100M_d22_seed2_2026-07-22_10-32-54_742661-pt

Trained with nanochat. Checkpoint at step 793.

W&B run: https://wandb.ai/alexksternteam/eval_seed_fleet_100M_v1/runs/lmmgxqwo

Outcome

metric	value
step	793
smooth_train_loss	2.9973487854003906
min_objective	1.0693600154337728
flops_used	9.525101529577554e+17
flops_per_token	4582014976.0
total_training_time	613.7771198749542

Training config

{
  "model_pt": {
    "sequence_len": 2048,
    "vocab_size": 65536,
    "n_layer": 22,
    "n_head": 11,
    "n_kv_head": 11,
    "n_embd": 1408
  },
  "model_ppt": {
    "sequence_len": 2048,
    "vocab_size": 1028,
    "n_layer": 22,
    "n_head": 11,
    "n_kv_head": 11,
    "n_embd": 1408
  },
  "optim": {
    "matrix_lr": 0.02,
    "embedding_lr": 0.3,
    "unembedding_lr": 0.004,
    "weight_decay": 0.0
  },
  "train": {
    "num_iterations": 1000,
    "eval_every_n_steps": null,
    "eval_every_flops_frac": null,
    "save_every_n_steps": null,
    "save_at_steps": null,
    "log_every_n_steps": 100,
    "eval_at_end": true,
    "save_at_end": true,
    "grad_clip": 1.0,
    "ema_beta": 0.0
  },
  "eval": {
    "eval_steps": 16,
    "eval_tokens": 2016000
  },
  "hardware": {
    "device_batch_size": 32,
    "grad_accum_steps": 2,
    "peak_tflops": 2250.0
  },
  "wandb": {
    "enabled": true,
    "notes": "",
    "group": "1e18_ppt0.05",
    "tags": [
      "nanochat_gpt",
      "pt_fineweb-nanochatbpe-100M",
      "ppt_declref-01-seq_len_2048-2B",
      "seed_2",
      "case_c",
      "lr_trapezoid",
      "depth_22"
    ],
    "entity": null
  },
  "data_pt": "fineweb-nanochatbpe-100M",
  "data_ppt": "declref-01-seq_len_2048-2B",
  "extra_eval": [
    "c4-nanochatbpe-10B"
  ],
  "train_split": "train",
  "pad_vocab": 65536,
  "ppt_vocab_size": null,
  "ppt_same_vocab_as_pt": false,
  "ppt_unified_lr": false,
  "lr_kind": "trapezoid",
  "lr_warmup_ratio": 0.0,
  "lr_warmdown_ratio": 0.4,
  "lr_final_frac": 0.0,
  "ppt_lr_kind": "trapezoid",
  "ppt_lr_warmup_ratio": 0.0,
  "ppt_lr_warmdown_ratio": 0.4,
  "ppt_lr_final_frac": 0.0,
  "ppt_lr": 0.006,
  "ppt_weight_decay": 0.0,
  "ppt_device_batch_size": 32,
  "ppt_grad_accum_steps": 8,
  "eval_tokens": 2016000,
  "reinit_embed_at_transition": true,
  "moment_match_embed_reinit": false,
  "reset_optimizer_at_transition": true,
  "random_pt_head": false,
  "depth": 22,
  "compile_model": true,
  "model_project": "eval_seed_fleet_100M_v1",
  "run_name": "declref_5pct_100M_d22_seed2",
  "target_flops": 1e+18,
  "alpha_ppt": 0.05,
  "use_measured_flops": true,
  "flops_per_token": null,
  "ppt_flops_per_token": null,
  "seed": 2,
  "metrics_output_file": null,
  "push_to_hf": true,
  "hf_repo_org": "alexkstern",
  "cleanup_checkpoints": true
}

Files

• model_000793.pt — model weights (state_dict).
• meta_000793.json — training metadata.
• config_000793.json — run config snapshot.
• rng_000793.pt — RNG state (when present).