code_5pct_3e18_seed2_2026-07-12_02-10-08_098372-ppt

Trained with nanochat. Checkpoint at step 275.

W&B run: https://wandb.ai/alexksternteam/optimal_seed_replicas_v1/runs/icxikmft

Outcome

metric	value
step	275
smooth_train_loss	2.775472640991211
min_objective	1.7841455652069105
flops_used	1.499733860286464e+17
flops_per_token	2080374784.0
total_training_time	51.19044303894043

Training config

{
  "model_pt": {
    "sequence_len": 2048,
    "vocab_size": 65536,
    "n_layer": 16,
    "n_head": 8,
    "n_kv_head": 8,
    "n_embd": 1024
  },
  "model_ppt": {
    "sequence_len": 2048,
    "vocab_size": 65536,
    "n_layer": 16,
    "n_head": 8,
    "n_kv_head": 8,
    "n_embd": 1024
  },
  "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": null,
    "save_every_n_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": "3e18_ppt0.05",
    "tags": [
      "nanochat_gpt",
      "pt_fineweb-nanochatbpe-20B",
      "ppt_github-code-nanochatbpe-1B",
      "seed_2",
      "case_b",
      "lr_trapezoid",
      "depth_16"
    ],
    "entity": null
  },
  "data_pt": "fineweb-nanochatbpe-20B",
  "data_ppt": "github-code-nanochatbpe-1B",
  "extra_eval": [
    "c4-nanochatbpe-10B"
  ],
  "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": 0.02,
  "lr_warmdown_ratio": 0.5,
  "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,
  "reset_optimizer_at_transition": false,
  "depth": 16,
  "compile_model": true,
  "model_project": "optimal_seed_replicas_v1",
  "run_name": "code_5pct_3e18_seed2",
  "target_flops": 3e+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_000275.pt — model weights (state_dict).
• meta_000275.json — training metadata.
• config_000275.json — run config snapshot.
• rng_000275.pt — RNG state (when present).