Views
No views yet
| metric | value |
|---|---|
step | 476 |
smooth_train_loss | 3.9976258277893066 |
min_objective | 5.792123248003767 |
flops_used | 2.563585027438805e+18 |
flops_per_token | 5136187392.0 |
total_training_time | 724.637948513031 |
1{
2 "model_pt": {
3 "sequence_len": 2048,
4 "vocab_size": 65536,
5 "n_layer": 24,
6 "n_head": 12,
7 "n_kv_head": 12,
8 "n_embd": 1536
9 },
10 "model_ppt": {
11 "sequence_len": 2048,
12 "vocab_size": 256,
13 "n_layer": 24,
14 "n_head": 12,
15 "n_kv_head": 12,
16 "n_embd": 1536
17 },
18 "optim": {
19 "matrix_lr": 0.01,
20 "embedding_lr": 0.01,
21 "unembedding_lr": 0.01,
22 "weight_decay": 0.0
23 },
24 "train": {
25 "num_iterations": 1000,
26 "eval_every_n_steps": 238,
27 "eval_every_flops_frac": null,
28 "save_every_n_steps": null,
29 "save_at_steps": [
30 17166,
31 19073,
32 36240,
33 38147,
34 55313,
35 57220,
36 74387,
37 76294,
38 95367
39 ],
40 "log_every_n_steps": 100,
41 "eval_at_end": true,
42 "save_at_end": true,
43 "grad_clip": 1.0,
44 "ema_beta": 0.0
45 },
46 "eval": {
47 "eval_steps": 8,
48 "eval_tokens": 2016000
49 },
50 "hardware": {
51 "device_batch_size": 32,
52 "grad_accum_steps": 4,
53 "peak_tflops": 2250.0
54 },
55 "wandb": {
56 "enabled": true,
57 "notes": "",
58 "group": "5e10tok_ppt0.01",
59 "tags": [
60 "nanochat_gpt",
61 "pt_fineweb-nanochatbpe-55B",
62 "ppt_dyck-k128-seq_len_2048-1B",
63 "seed_0",
64 "case_c",
65 "lr_trapezoid",
66 "depth_24"
67 ],
68 "entity": null
69 },
70 "data_pt": "fineweb-nanochatbpe-55B",
71 "data_ppt": "dyck-k128-seq_len_2048-1B",
72 "extra_eval": [
73 "c4-nanochatbpe-10B"
74 ],
75 "train_split": "train",
76 "pad_vocab": 65536,
77 "ppt_vocab_size": null,
78 "ppt_same_vocab_as_pt": false,
79 "ppt_unified_lr": false,
80 "lr_kind": "trapezoid",
81 "lr_warmup_ratio": 0.1,
82 "lr_warmdown_ratio": 0.4,
83 "lr_final_frac": 0.0,
84 "ppt_lr_kind": "trapezoid",
85 "ppt_lr_warmup_ratio": 0.1,
86 "ppt_lr_warmdown_ratio": 0.4,
87 "ppt_lr_final_frac": 0.0,
88 "ppt_lr": 0.01,
89 "ppt_weight_decay": 0.0,
90 "ppt_device_batch_size": 32,
91 "ppt_grad_accum_steps": 4,
92 "eval_tokens": 2016000,
93 "reinit_embed_at_transition": true,
94 "moment_match_embed_reinit": false,
95 "reset_optimizer_at_transition": true,
96 "random_pt_head": false,
97 "depth": 24,
98 "compile_model": true,
99 "model_project": "longhorizon_50B_v0",
100 "run_name": "kdyck_1pct_50B_d24_seed0_lr0p0036813",
101 "target_flops": 0.0,
102 "ppt_tokens": 500000000,
103 "pt_tokens": 49500000000,
104 "ppt_eval_every_n_steps": 238,
105 "alpha_ppt": 0.0,
106 "use_measured_flops": true,
107 "flops_per_token": null,
108 "ppt_flops_per_token": null,
109 "seed": 0,
110 "metrics_output_file": null,
111 "push_to_hf": true,
112 "hf_repo_org": "alexkstern",
113 "cleanup_checkpoints": false
114}model_000476.pt — model weights (state_dict).meta_000476.json — training metadata.config_000476.json — run config snapshot.rng_000476.pt — RNG state (when present).