letterpress / large /context.md
disco-jack-basement's picture
provenance: large/context.md
896b90d verified
|
Raw
History Blame Contribute Delete
1.85 kB

large v1

  • preset: large
  • training corpus: data/large (see corpus_index.md / corpus_stats.txt)
  • trained: 2026-06-18 13:58 (checkpoint mtime)
  • training wall-clock: 208.2 min
  • final val loss: 1.1111 (iter 23000)
  • parameters: 49.72M | vocab 198
  • note: v1 trained on the large corpus as of 2026-06-18 BEFORE the xlarge expansion: 1,085,579,228 chars / 248 authors / 22 categories / vocab 198 / best val 1.1111. The corpus_index.md + corpus_stats.txt in this dir have been CORRECTED to that true training set (the live data/large has since grown toward the xlarge ~2B target).

Training hyperparameters

param value
block_size 384
n_embd 640
n_head 10
n_layer 10
dropout 0.0
batch_size 16
max_iters 24000
eval_interval 1000
eval_iters 80
learning_rate 0.00025
vocab_size 198

Validation curve

step train val
0 5.5531 5.5478
1000 2.0925 2.0524
2000 1.5805 1.5329
3000 1.4392 1.3825
4000 1.3836 1.3366
5000 1.3361 1.2880
6000 1.3129 1.2693
7000 1.2897 1.2436
8000 1.2580 1.2296
9000 1.2388 1.2164
10000 1.2397 1.1977
11000 1.2146 1.1828
12000 1.2159 1.1787
13000 1.2004 1.1669
14000 1.1967 1.1636
15000 1.1915 1.1576
16000 1.1814 1.1457
17000 1.1793 1.1388
18000 1.1634 1.1461
19000 1.1557 1.1360
20000 1.1552 1.1365
21000 1.1573 1.1241
22000 1.1556 1.1260
23000 1.1388 1.1111
23999 1.1367 1.1189

Reproducing the corpus

corpus_index.md lists every author and work in the training set; re-run python -m corpus add-author / add-topic per that index and make finalize to rebuild an equivalent corpus.