Views
No views yet
| metric | value |
|---|---|
step | 5004 |
smooth_train_loss | 4.659100081066534 |
min_objective | 1.47160866154512 |
flops_used | 5.183874448364667e+17 |
flops_per_token | 3161456704.0 |
total_training_time | 835.4741947650909 |
1{
2 "model_pt": {
3 "sequence_len": 1024,
4 "vocab_size": 64000,
5 "n_layer": 24,
6 "n_head": 16,
7 "n_embd": 1024,
8 "intermediate_size": 4096,
9 "rope_theta": 10000.0,
10 "attention_dropout": 0.1,
11 "weight_tying": true,
12 "head_bias": true,
13 "initializer_range": 0.02
14 },
15 "model_ppt": null,
16 "optim": {
17 "matrix_lr": 0.0001,
18 "embedding_lr": 0.0001,
19 "unembedding_lr": 0.0001,
20 "weight_decay": 0.0
21 },
22 "train": {
23 "num_iterations": 1000,
24 "eval_every_n_steps": 1000,
25 "save_every_n_steps": null,
26 "log_every_n_steps": 1,
27 "eval_at_end": true,
28 "save_at_end": true,
29 "grad_clip": 0.0
30 },
31 "eval": {
32 "eval_steps": 62,
33 "eval_tokens": 2016000
34 },
35 "hardware": {
36 "device_batch_size": 8,
37 "grad_accum_steps": 1,
38 "peak_tflops": 2250.0
39 },
40 "wandb": {
41 "enabled": true,
42 "notes": "",
43 "group": "repro_c4_owt",
44 "tags": [
45 "hf_llama",
46 "vocab_64000",
47 "pt_owt-gpt2bpe-9B",
48 "seed_42",
49 "ppt_c4-gpt2bpe-10B",
50 "ppt_lr_cosine",
51 "reinit"
52 ],
53 "entity": null
54 },
55 "data_pt": "owt-gpt2bpe-9B",
56 "data_ppt": "c4-gpt2bpe-10B",
57 "extra_eval": [
58 "fineweb-gpt2bpe-20B",
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": 164000000,
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": false,
82 "model_project": "nca-repro",
83 "run_name": "c4_owt",
84 "target_flops": 0.0,
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_005004.pt — model weights (state_dict).meta_005004.json — training metadata.config_005004.json — run config snapshot.rng_005004.pt — RNG state (when present).