Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,139 @@
+---
+library_name: nanochat
+license: apache-2.0
+tags:
+- nanochat
+- hf_llama
+- vocab_64000
+- pt_owt-gpt2bpe-9B
+- seed_0
+- ppt_c4-gpt2bpe-10B
+- ppt_lr_cosine
+- reinit
+---
+
+# c4_owt_untied_seed0_2026-07-20_18-36-40_388310-owt
+
+Trained with [nanochat](https://github.com/karpathy/nanochat). Checkpoint at step **17,196**.
+
+**W&B run:** https://wandb.ai/alexksternteam/nca-repro/runs/sewbgrll
+
+## Outcome
+
+| metric | value |
+| --- | --- |
+| `step` | 17196 |
+| `smooth_train_loss` | 2.8550471171297045 |
+| `min_objective` | 0.9622754070976698 |
+| `flops_used` | 2.8502607518490427e+19 |
+| `flops_per_token` | 3161456704.0 |
+| `total_training_time` | 58587.57112932205 |
+
+## Training config
+
+```json
+{
+  "model_pt": {
+    "sequence_len": 1024,
+    "vocab_size": 64000,
+    "n_layer": 24,
+    "n_head": 16,
+    "n_embd": 1024,
+    "intermediate_size": 4096,
+    "rope_theta": 10000.0,
+    "attention_dropout": 0.1,
+    "weight_tying": false,
+    "head_bias": false,
+    "initializer_range": 0.02,
+    "nanochat_init": false,
+    "attn_implementation": "sdpa",
+    "explicit_additive_mask": true
+  },
+  "model_ppt": null,
+  "optim": {
+    "matrix_lr": 0.0005,
+    "embedding_lr": 0.0005,
+    "unembedding_lr": 0.0005,
+    "weight_decay": 0.0001
+  },
+  "train": {
+    "num_iterations": 1000,
+    "eval_every_n_steps": 1000,
+    "eval_every_flops_frac": null,
+    "save_every_n_steps": null,
+    "save_at_steps": null,
+    "log_every_n_steps": 10,
+    "eval_at_end": true,
+    "save_at_end": true,
+    "grad_clip": 1.0,
+    "ema_beta": 0.9
+  },
+  "eval": {
+    "eval_steps": 16,
+    "eval_tokens": 2016000
+  },
+  "hardware": {
+    "device_batch_size": 64,
+    "grad_accum_steps": 4,
+    "peak_tflops": 2250.0
+  },
+  "wandb": {
+    "enabled": true,
+    "notes": "",
+    "group": "repro_c4_owt",
+    "tags": [
+      "hf_llama",
+      "vocab_64000",
+      "pt_owt-gpt2bpe-9B",
+      "seed_0",
+      "ppt_c4-gpt2bpe-10B",
+      "ppt_lr_cosine",
+      "reinit"
+    ],
+    "entity": null
+  },
+  "data_pt": "owt-gpt2bpe-9B",
+  "data_ppt": "c4-gpt2bpe-10B",
+  "extra_eval": [
+    "fineweb-gpt2bpe-20B",
+    "nca-paper-1024"
+  ],
+  "train_split": "train",
+  "pad_vocab": 64000,
+  "lr_kind": "cosine",
+  "lr_warmup_ratio": 0.1,
+  "lr_warmdown_ratio": 0.0,
+  "lr_final_frac": 0.0,
+  "ppt_lr": 0.0001,
+  "ppt_weight_decay": 0.0,
+  "ppt_grad_clip": 0.0,
+  "ppt_device_batch_size": 8,
+  "ppt_grad_accum_steps": 2,
+  "ppt_lr_kind": "cosine",
+  "ppt_lr_warmup_ratio": 0.1,
+  "ppt_lr_warmdown_ratio": 0.0,
+  "ppt_tokens": 164000000,
+  "pt_tokens": null,
+  "eval_tokens": 2016000,
+  "reinit_embed_at_transition": true,
+  "reset_optimizer_at_transition": true,
+  "depth": null,
+  "compile_model": false,
+  "model_project": "nca-repro",
+  "run_name": "c4_owt_untied_seed0",
+  "target_flops": 0.0,
+  "alpha_ppt": 0.0,
+  "use_measured_flops": true,
+  "flops_per_token": null,
+  "seed": 0,
+  "push_to_hf": true,
+  "hf_repo_org": "alexkstern"
+}
+```
+
+## Files
+
+- `model_017196.pt` — model weights (`state_dict`).
+- `meta_017196.json` — training metadata.
+- `config_017196.json` — run config snapshot.
+- `rng_017196.pt` — RNG state (when present).