Merge branch 'master' into fix/typo

2026-03-15 22:18:01 +01:00
parent 6405b26d24 1b1cc3c599
commit 02e865c2ab
6 changed files with 78 additions and 18 deletions
@@ -196,3 +196,7 @@ NOTE: The `val_bpb` is as of this run *NOT* comparable due to the data distribut
 Achieved Mar 9, 2026 on commit `6ed7d1d`. Exactly the same launch command as Run 4 except `--target-param-data-ratio=8.7`. I ran 5 identical runs, the average CORE was 0.2690, which is quite a bit above the needed threshold of 0.2565. But the reason I didn't decrease the ratio further (i.e. train shorter) is that while the CORE "safety gap" is large, the val_loss safety gap is smaller - 0.71808, which we want to be below the Run 4 val loss of 0.71854. It's likely that we could have reduced the ratio even lower, possibly to 8.6, but it's not worth splitting hairs at this point.

 This commit is special because all of the improvements that went into [this commit](https://github.com/karpathy/nanochat/commit/6ed7d1d82cee16c2e26f45d559ad3338447a6c1b) came from fully autonomous "research" done by a private version of [autoresearch](https://github.com/karpathy/autoresearch) run on a d12 model. I wrote more about this in [this tweet](https://x.com/karpathy/status/2031135152349524125). The changes easily translated from d12 to d24, hence new leaderboard record, taking us from 2.02 hours "time to GPT-2" to 1.80 hours.
+
+## Run 6
+
+Achieved Mar 14, 2026 on commit `a825e63`. Exactly the same launch command as Run 4 except `--target-param-data-ratio=8`. Improvements in the architecture are allowing us to train shorter and shorter time. Instead of an undertrained d24 I attempted to train an overtrained d22 but it was worse. This set of changes came from autoresearch round 2, where I asked it to reference the modded-nanogpt repo for inspiration. So the exploration tried out a number of ideas and in particular found a way to incorporate the backout and smear in such a way that they are helpful (I had previously tried them manually a long time ago and they caused regressions). The smear idea in particular is a little bit heavier and bloaty because it is essentially an "early fusion" of context across tokens, producing a kind of a bigram input into the network and allowing it to focus on higher ngrams earlier. But for this reason the code gets a bit more complex and required some changes to inference. I verified with a unit test that the Engine inference is correct compared to the naive inference of `GPT.generate()`. The average of 5 runs was CORE 0.262634 and each of them lasted 1.65 hours (99 minutes).