Views
No views yet
| metric | value |
|---|---|
step | 28859 |
smooth_train_loss | 2.8787535675315685 |
min_objective | 0.9639769115633504 |
flops_used | 2.9021789466451247e+19 |
flops_per_token | 1918107712.0 |
total_training_time | 23852.965278863907 |
1{
2 "model_pt": {
3 "sequence_len": 1024,
4 "vocab_size": 64000,
5 "n_layer": 24,
6 "n_head": 12,
7 "n_embd": 768,
8 "intermediate_size": 3072,
9 "rope_theta": 10000.0,
10 "attention_dropout": 0.1,
11 "weight_tying": true,
12 "head_bias": true,
13 "initializer_range": 0.02,
14 "attn_implementation": "sdpa"
15 },
16 "model_ppt": null,
17 "optim": {
18 "matrix_lr": 0.0005,
19 "embedding_lr": 0.0005,
20 "unembedding_lr": 0.0005,
21 "weight_decay": 0.0001
22 },
23 "train": {
24 "num_iterations": 1000,
25 "eval_every_n_steps": 1000,
26 "save_every_n_steps": null,
27 "log_every_n_steps": 10,
28 "eval_at_end": true,
29 "save_at_end": true,
30 "grad_clip": 1.0
31 },
32 "eval": {
33 "eval_steps": 4,
34 "eval_tokens": 2016000
35 },
36 "hardware": {
37 "device_batch_size": 128,
38 "grad_accum_steps": 1,
39 "peak_tflops": 2250.0
40 },
41 "wandb": {
42 "enabled": true,
43 "notes": "",
44 "group": "repro_vanilla_owt",
45 "tags": [
46 "hf_llama",
47 "vocab_64000",
48 "pt_owt-gpt2bpe-9B",
49 "seed_42",
50 "no_ppt"
51 ],
52 "entity": null
53 },
54 "data_pt": "owt-gpt2bpe-9B",
55 "data_ppt": null,
56 "extra_eval": [
57 "fineweb-gpt2bpe-20B",
58 "c4-gpt2bpe-10B",
59 "nca-paper-1024"
60 ],
61 "train_split": "train",
62 "pad_vocab": 64000,
63 "lr_kind": "cosine",
64 "lr_warmup_ratio": 0.1,
65 "lr_warmdown_ratio": 0.0,
66 "lr_final_frac": 0.0,
67 "ppt_lr": 0.0001,
68 "ppt_weight_decay": 0.0,
69 "ppt_grad_clip": 0.0,
70 "ppt_device_batch_size": 8,
71 "ppt_grad_accum_steps": 1,
72 "ppt_lr_kind": "cosine",
73 "ppt_lr_warmup_ratio": 0.1,
74 "ppt_lr_warmdown_ratio": 0.0,
75 "ppt_tokens": null,
76 "pt_tokens": null,
77 "eval_tokens": 2016000,
78 "reinit_embed_at_transition": true,
79 "reset_optimizer_at_transition": true,
80 "depth": null,
81 "compile_model": true,
82 "model_project": "nca-repro",
83 "run_name": "vanilla_owt_270M",
84 "target_flops": 2.9022e+19,
85 "alpha_ppt": 0.0,
86 "use_measured_flops": true,
87 "seed": 42,
88 "push_to_hf": true,
89 "hf_repo_org": "alexkstern"
90}model_028859.pt — model weights (state_dict).meta_028859.json — training metadata.config_028859.json — run config snapshot.rng_028859.pt — RNG state (when present).