feat: nanochat d12 step 87k (286M params, val_bpb=0.866)
Browse files- README.md +16 -0
- metadata.json +61 -0
- model.pt +3 -0
README.md
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
license: mit
|
| 4 |
+
library_name: pytorch
|
| 5 |
+
tags: [nanochat, gpt, pretraining]
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
# nanochat-d12-step87k
|
| 9 |
+
|
| 10 |
+
286M parameter GPT model trained with nanochat framework.
|
| 11 |
+
|
| 12 |
+
- **Architecture**: 12 layers, 768 dim, 6 heads, RoPE, GQA, ReLU² MLP
|
| 13 |
+
- **Context**: 2048 tokens, full attention (window_pattern=L)
|
| 14 |
+
- **Training**: 87,000 steps, ~5.7B tokens, Chinchilla-optimal (ratio=12)
|
| 15 |
+
- **Val bpb**: 0.8658
|
| 16 |
+
- **GPU**: RTX 4070 12GB, bf16, 28.4 hours
|
metadata.json
ADDED
|
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"step": 87000,
|
| 3 |
+
"val_bpb": 0.8657516205114915,
|
| 4 |
+
"model_config": {
|
| 5 |
+
"sequence_len": 2048,
|
| 6 |
+
"vocab_size": 32768,
|
| 7 |
+
"n_layer": 12,
|
| 8 |
+
"n_head": 6,
|
| 9 |
+
"n_kv_head": 6,
|
| 10 |
+
"n_embd": 768,
|
| 11 |
+
"window_pattern": "L"
|
| 12 |
+
},
|
| 13 |
+
"user_config": {
|
| 14 |
+
"run": "rtx4070-d12-chinchilla",
|
| 15 |
+
"tracker": "wandb",
|
| 16 |
+
"mlflow_uri": null,
|
| 17 |
+
"mlflow_experiment": "nanochat",
|
| 18 |
+
"device_type": "",
|
| 19 |
+
"fp8": false,
|
| 20 |
+
"fp8_recipe": "tensorwise",
|
| 21 |
+
"depth": 12,
|
| 22 |
+
"aspect_ratio": 64,
|
| 23 |
+
"head_dim": 128,
|
| 24 |
+
"max_seq_len": 2048,
|
| 25 |
+
"window_pattern": "L",
|
| 26 |
+
"num_iterations": 87000,
|
| 27 |
+
"target_flops": -1.0,
|
| 28 |
+
"target_param_data_ratio": 12,
|
| 29 |
+
"device_batch_size": 8,
|
| 30 |
+
"total_batch_size": 65536,
|
| 31 |
+
"embedding_lr": 0.3,
|
| 32 |
+
"unembedding_lr": 0.008,
|
| 33 |
+
"weight_decay": 0.28,
|
| 34 |
+
"matrix_lr": 0.02,
|
| 35 |
+
"scalar_lr": 0.5,
|
| 36 |
+
"warmup_steps": 40,
|
| 37 |
+
"warmdown_ratio": 0.65,
|
| 38 |
+
"final_lr_frac": 0.05,
|
| 39 |
+
"resume_from_step": 10000,
|
| 40 |
+
"eval_every": 2000,
|
| 41 |
+
"eval_tokens": 524288,
|
| 42 |
+
"core_metric_every": 10000,
|
| 43 |
+
"core_metric_max_per_task": 200,
|
| 44 |
+
"sample_every": 5000,
|
| 45 |
+
"save_every": 10000,
|
| 46 |
+
"model_tag": null
|
| 47 |
+
},
|
| 48 |
+
"device_batch_size": 8,
|
| 49 |
+
"max_seq_len": 2048,
|
| 50 |
+
"total_batch_size": 65536,
|
| 51 |
+
"dataloader_state_dict": {
|
| 52 |
+
"pq_idx": 10,
|
| 53 |
+
"rg_idx": 77,
|
| 54 |
+
"epoch": 1
|
| 55 |
+
},
|
| 56 |
+
"loop_state": {
|
| 57 |
+
"min_val_bpb": 0.8657516205114915,
|
| 58 |
+
"smooth_train_loss": 2.7484221461371146,
|
| 59 |
+
"total_training_time": 102370.698939085
|
| 60 |
+
}
|
| 61 |
+
}
|
model.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:95d2ef6e9d97e135b8947f8ec3fd1d37bac1b43a49a471f3096811265cd27b38
|
| 3 |
+
size 792761399
|