c4_5pct_27e18_d22_seed1_2026-08-03_11-51-28_317697-ppt

Trained with nanochat. Checkpoint at step 1,123.

W&B run: https://wandb.ai/alexksternteam/evalcurve_27e18_v0/runs/35zupcl8

Outcome

metric	value
step	1123
smooth_train_loss	3.7412948608398438
min_objective	1.1194889340805514
flops_used	1.348888905134375e+18
flops_per_token	4582014976.0
total_training_time	435.7113516330719

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": 65536,
    "n_layer": 22,
    "n_head": 11,
    "n_kv_head": 11,
    "n_embd": 1408
  },
  "optim": {
    "matrix_lr": 0.015,
    "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": 0.05,
    "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": 8,
    "eval_tokens": 2016000
  },
  "hardware": {
    "device_batch_size": 32,
    "grad_accum_steps": 1,
    "peak_tflops": 2250.0
  },
  "wandb": {
    "enabled": true,
    "notes": "",
    "group": "3e19_ppt0.05",
    "tags": [
      "nanochat_gpt",
      "pt_fineweb-nanochatbpe-20B",
      "ppt_c4-nanochatbpe-10B",
      "seed_1",
      "case_b",
      "lr_trapezoid",
      "depth_22"
    ],
    "entity": null
  },
  "data_pt": "fineweb-nanochatbpe-20B",
  "data_ppt": "c4-nanochatbpe-10B",
  "extra_eval": [
    "github-code-nanochatbpe-1B"
  ],
  "train_split": "train",
  "pad_vocab": 65536,
  "ppt_vocab_size": null,
  "ppt_same_vocab_as_pt": true,
  "ppt_unified_lr": true,
  "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": null,
  "ppt_lr_warmdown_ratio": null,
  "ppt_lr_final_frac": 0.0,
  "ppt_lr": null,
  "ppt_weight_decay": 0.0,
  "ppt_device_batch_size": null,
  "ppt_grad_accum_steps": null,
  "eval_tokens": 2016000,
  "reinit_embed_at_transition": false,
  "moment_match_embed_reinit": false,
  "reset_optimizer_at_transition": false,
  "random_pt_head": false,
  "depth": 22,
  "compile_model": true,
  "model_project": "evalcurve_27e18_v0",
  "run_name": "c4_5pct_27e18_d22_seed1",
  "target_flops": 2.7e+19,
  "ppt_tokens": null,
  "pt_tokens": null,
  "ppt_eval_every_n_steps": null,
  "alpha_ppt": 0.05,
  "use_measured_flops": true,
  "flops_per_token": null,
  "ppt_flops_per_token": null,
  "seed": 1,
  "metrics_output_file": null,
  "push_to_hf": true,
  "hf_repo_org": "alexkstern",
  "cleanup_checkpoints": true
}

Files

• model_001123.pt — model weights (state_dict).
• meta_001123.json — training metadata.
• config_001123.json — run config snapshot.
• rng_001123.pt — RNG state (when present).