78 Commits
Author SHA1 Message Date
wassnameandPI[openai-codex] 4ff0710851 Score choices after the answer newline
Co-Authored-By: PI[openai-codex] <288921227+claudypoo@users.noreply.github.com>
2026-09-02 18:42:28 +08:00
wassnameandPI[openai-codex] 88846c19a7 Audit paper-v5 Gemma-1B seed 42 run
Co-Authored-By: PI[openai-codex] <288921227+claudypoo@users.noreply.github.com>
2026-09-02 16:46:28 +08:00
wassnameandPI[openai-codex] d8bea05819 Credit Foomax for replication testing
Co-Authored-By: PI[openai-codex] <288921227+claudypoo@users.noreply.github.com>
2026-09-02 14:08:38 +08:00
wassnameandPI[openai-codex] 6624f368b8 Make paper training presets explicit
Co-Authored-By: PI[openai-codex] <288921227+claudypoo@users.noreply.github.com>
2026-09-02 13:26:43 +08:00
wassnameandClaude Opus 5 5e0f8517f3 Correct the results table to the v5 paper numbers
The table I added in a3e30ac came from a January draft (14.1 F1, 3.1x) and
contradicted the README's own 6.9x. arXiv v5 and main.tex give 31.2+-5.3
against 4.5 for prompting.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-21 14:00:27 +08:00
wassnameandClaude Opus 5 a3e30aceaa Add the DailyDilemmas results table and the CoLoRAI acceptance to the README
The headline numbers only lived in the arXiv PDF, so anything scraping the
repo could not see them.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-21 13:45:14 +08:00
wassname 0202660e94 update justfile 2026-02-22 09:35:47 +08:00
wassname 3747447fee asym_coh_ratio 2026-02-21 21:03:35 +08:00
wassname c1575003e0 smaller init 2026-02-21 21:02:40 +08:00
wassname 77c9634bcf fixes 2026-02-21 21:02:22 +08:00
wassname 85c09b77fe sp 2026-02-08 16:39:31 +08:00
wassname 2aff548a7a add tests 2026-02-08 16:35:23 +08:00
wassname bf3f8faad2 update title 2026-02-08 13:52:07 +08:00
wassname (Michael J Clark) 85e1f9aab6 Update README.md 2026-02-07 11:54:08 +08:00
wassname (Michael J Clark) 2dfdfc6328 Update README.md 2026-02-07 11:53:33 +08:00
wassname (Michael J Clark) ffb2087fdf Update README.md 2026-02-07 11:51:31 +08:00
wassname (Michael J Clark) d40078a407 Update README.md 2026-02-07 11:48:07 +08:00
wassname c36b3ad4d6 Merge branch 'main' of https://github.com/wassname/AntiPaSTO 2026-01-18 12:36:30 +08:00
wassname f5a1ed8bbb lower pmass thresh 2026-01-18 12:36:28 +08:00
wassname (Michael J Clark) 251dffeee4 Update README.md 2026-01-18 08:53:20 +08:00
wassname (Michael J Clark) 0475e80b3c Update README.md 2026-01-18 08:51:33 +08:00
wassname 1b85f7f808 Merge branch 'main' of https://github.com/wassname/AntiPaSTO 2026-01-17 20:36:27 +08:00
wassname e65fd34222 fix format 2026-01-17 20:36:10 +08:00
wassname b4a35bb724 Merge branch 'main' of https://github.com/wassname/AntiPaSTO 2026-01-16 11:44:00 +00:00
wassname ddcd830a4b gradual warmup 2026-01-16 09:47:35 +00:00
wassname (Michael J Clark) 081ca634c2 Update README.md by modifying image placement
Removed an image of the loss landscape and added it back later in the document.
2026-01-16 15:23:35 +08:00
wassname (Michael J Clark) 226fd408b8 Replace image with new loss landscape visualization
Updated the image in the README to illustrate the loss landscape with a new visual representation.
2026-01-16 15:18:32 +08:00
wassname (Michael J Clark) 43bca49942 Update README.md 2026-01-16 15:17:21 +08:00
wassname 08d54c7095 bug fix 2026-01-16 06:09:28 +00:00
wassname 4de4166d4e refine approx_intersection_bases: increase min_overlap to 0.5 and add diagnostic logging 2026-01-16 02:57:19 +00:00
wassname fa2d74649b Merge branch 'main' of https://github.com/wassname/AntiPaSTO 2026-01-16 02:55:57 +00:00
wassname 06ac1b7e1a early stop after all warmups 2026-01-16 02:16:33 +00:00
wassname 9e6732d71d demo 2026-01-15 19:32:20 +08:00
wassname 5cb8c31273 demo 2026-01-15 16:27:52 +08:00
wassname b2985b90eb fix metric print 2026-01-15 14:04:59 +08:00
wassname 38546c03d1 add val loss to info 2026-01-15 12:28:24 +08:00
wassname b05cf9aad2 fix bug in focus calc 2026-01-15 12:27:45 +08:00
wassname f253e07742 name ckpt 2026-01-15 12:11:42 +08:00
wassname ad203ebb9a loss display 2026-01-15 12:11:37 +08:00
wassname b29addb24d compute focus in raw space, improving numerical stability 2026-01-15 11:53:28 +08:00
wassname acaa60d556 tidy talk to checkpoint notebook 2026-01-15 11:52:57 +08:00
wassname c98a5a239a add & simplify warmup for mono, coh, and now focus 2026-01-15 11:52:38 +08:00
wassname 4b316f038f computes focus in raw space (not fisher which leads to invalid ratios) 2026-01-14 18:26:23 +08:00
wassname aadc367e6e Merge branch 'main' of https://github.com/wassname/AntiPaSTO 2026-01-14 15:44:38 +08:00
wassname 83e16ebc2e demo load from hf! 2026-01-14 15:43:38 +08:00
wassname f3990e467e talk 2026-01-14 15:39:39 +08:00
wassname e5b150fe23 Merge branch 'main' of https://github.com/wassname/AntiPaSTO 2026-01-14 02:06:43 +00:00
wassname 32115c3608 simplify loss, Training complete. Final loss: -6.1250 2026-01-14 02:06:40 +00:00
wassname 1e4236ebf0 Merge branch 'main' of https://github.com/wassname/AntiPaSTO 2026-01-14 09:16:15 +08:00
wassname d8a3f52b16 shouldn't have symlog around this loss, simplify 2026-01-14 00:50:11 +00:00