Andrej Karpathy 6ed7d1d82c All of these improvements were developed by Claude running autonomously over ~2 days using autoresearch. I didn't touch anything - incredible. All tuning was done on d12 but generalized easily to larger models (e.g. d24 in particular). This means we will also get a new "Time to GPT-2" Leaderboard entry, which I will push separately. ... Optimizer & schedule changes: - Increase unembedding LR 0.004 -> 0.008, weight decay 0.2 -> 0.28 - Per-group Adam betas and weight decay (instead of shared global betas) - Muon beta2 0.95 -> 0.9, momentum warmup target 0.95 -> 0.97 over 400 steps - Warmup: ratio-based -> absolute steps (default 40) - Warmdown ratio 0.5 -> 0.65, final LR fraction 0.0 -> 0.05 - Weight decay schedule: linear -> cosine decay - Polar express norm factor 1.02 -> 1.01 Architecture & init changes: - VE gate: channels 32 -> 12, scale range 2x -> 3x, init small positive - Add post-QK-norm scaling (q,k *= 1.15) for sharper attention - Embedding init std 1.0 -> 0.8, MLP c_fc init 0.5x smaller - RoPE base theta 10K -> 100K - Short attention window: seq_len/2 -> ~seq_len/3 (ceil to 128 tile) - Logit softcap 20 -> 15		2026-03-09 20:45:17 +00:00
..
__init__.py	initial commit	2025-10-13 06:49:24 -07:00
checkpoint_manager.py	tune the data mixture a bit, load optimizer by default when SFT. These were confirmed to be best settings from sweeps of sft	2026-02-18 15:49:18 +00:00
common.py	delete autocast, an unnecessary thorn in my side, manage dtypes directly	2026-03-04 23:55:30 +00:00
core_eval.py	initial commit	2025-10-13 06:49:24 -07:00
dataloader.py	big, breaking change but large upside: swap previous FineWeb-EDU dataset to NVIDIA ClimbMix dataset. Requires people to download the data shards. The upside is that training GPT-2 capablity model now only takes ~2 hours, down from 2.76 hours, so this is a huge win data-wise	2026-03-04 19:47:12 +00:00
dataset.py	big, breaking change but large upside: swap previous FineWeb-EDU dataset to NVIDIA ClimbMix dataset. Requires people to download the data shards. The upside is that training GPT-2 capablity model now only takes ~2 hours, down from 2.76 hours, so this is a huge win data-wise	2026-03-04 19:47:12 +00:00
engine.py	delete autocast, an unnecessary thorn in my side, manage dtypes directly	2026-03-04 23:55:30 +00:00
execution.py	nit delete redundant catch/raise in execute	2025-10-29 08:10:03 -07:00
flash_attention.py	delete autocast, an unnecessary thorn in my side, manage dtypes directly	2026-03-04 23:55:30 +00:00
fp8.py	delete autocast, an unnecessary thorn in my side, manage dtypes directly	2026-03-04 23:55:30 +00:00
gpt.py	All of these improvements were developed by Claude running autonomously over ~2 days using autoresearch. I didn't touch anything - incredible. All tuning was done on d12 but generalized easily to larger models (e.g. d24 in particular). This means we will also get a new "Time to GPT-2" Leaderboard entry, which I will push separately.	2026-03-09 20:45:17 +00:00
logo.svg	initial commit	2025-10-13 06:49:24 -07:00
loss_eval.py	fix typos	2025-11-14 11:20:25 +01:00
optim.py	All of these improvements were developed by Claude running autonomously over ~2 days using autoresearch. I didn't touch anything - incredible. All tuning was done on d12 but generalized easily to larger models (e.g. d24 in particular). This means we will also get a new "Time to GPT-2" Leaderboard entry, which I will push separately.	2026-03-09 20:45:17 +00:00
report.py	remove leftover mid references (#491)	2026-02-02 08:33:46 -08:00
tokenizer.py	adjust the comment on the regex pattern per recent experimnet see dev/LOG.md	2026-01-13 17:50:39 +00:00
ui.html	Fix conversation scroll to bottom on some browsers + remove duplicated padding (#348)	2025-12-31 13:03:22 -08:00