python-sandbox-math-demo / CHANGELOG.md
austindixson's picture
Oct 30 full-context results; Oct 28 accuracy gains labeled as cap effects (#1)
5d18f7c
|
Raw History Blame Contribute Delete
581 Bytes

Changelog

2026-10-30

  • RECORDED.md and the Recorded results tab: add the Oct 30 full-context results (BF16 LFM3-1.7B, A100, 262k context). The tool gave about the same accuracy as no tools and fewer tokens per solved problem. The Oct 28 accuracy gains are now labeled as cap effects. Notes that the cloud runs had network access (NET_OK) and did not use the network-denied sandbox.
  • index.html: show fig_oct30.png instead of the Oct 28 AIME accuracy chart. fig_aime.png stays in the repo.
  • README.md: one line on the full-context result with a link to the write-up.