wassname/jsteer
1
0
Fork 0
mirror of https://github.com/wassname/jsteer.git synced 2026-08-21 11:16:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Commit Graph

  • 6836cf5783 Dedup marimo-pair/retro-marimo-pair into home skills main wassnameandClaudypoo 2026-07-14 13:55:25 +08:00
  • 35230357bb Sync marimo-pair provenance header with home skill copy wassnameandClaudypoo 2026-07-14 13:25:30 +08:00
  • 5670cef162 Resolve model commit from cached config wassname 2026-07-13 06:56:52 +08:00
  • b927dc64fd Use the shared full-precision edge evaluator wassname 2026-07-13 06:50:02 +08:00
  • 1c927e95ba Keep demo reproduction on the fitted layer band wassname 2026-07-13 06:26:35 +08:00
  • 77998d6bad Make steering demo runs selectable and append-only wassname 2026-07-13 06:07:44 +08:00
  • 89fc7b1068 wip wassname 2026-07-13 05:38:29 +08:00
  • 22613ce40d nbs/steering_demo: fix persona_soft ans_mass mislabel in qualitative cell wassnameandClaudypoo 2026-07-12 16:33:28 +08:00
  • a428af4413 docs: oracle (deepseek-v4-pro) review + triage on why steering fails the verdict wassnameandClaudypoo 2026-07-12 16:29:46 +08:00
  • 74e21552af nbs/steering_demo: dual-gate results + sig-fig table + Claude qualitative read wassnameandClaudypoo 2026-07-12 16:29:46 +08:00
  • e0942af2e1 nbs: remove superseded persona_steering{,_v2,_v3} + demo_edges scratch wassnameandClaudypoo 2026-07-12 15:41:16 +08:00
  • 8cf966125d demo: coherent_edge dual gate min(rep, ans_mass) -- task-42 rep-only over-steered wassnameandClaudypoo 2026-07-12 15:41:07 +08:00
  • fb7f214d39 AGENTS: task-42 result -- rep-only over-steers, ans_mass binds first for verdict readout wassnameandClaudypoo 2026-07-12 14:41:40 +08:00
  • 0c02af1ba4 demo: raise max_C to 1e5 (safety only, edge set by rep); AGENTS: record C-scale problem wassnameandClaudypoo 2026-07-12 14:24:54 +08:00
  • 70c027ca46 demo: add score (validity-weighted swing) + at_budget flag to comparison table wassnameandClaudypoo 2026-07-12 14:23:02 +08:00
  • 349c9af3fe demo: calibrate coherent-edge on rep alone; ans_mass -> base-anchored readout flag wassnameandClaudypoo 2026-07-12 14:12:21 +08:00
  • 541fc49d4b AGENTS: correct calibration note -- ans_mass pre-empts rep, not rep insensitivity wassnameandClaudypoo 2026-07-12 14:02:59 +08:00
  • d6fe8c496b AGENTS: note steering-demo calibration is not method-comparable yet (open) wassnameandClaudypoo 2026-07-12 13:42:14 +08:00
  • ed7fa92996 nbs: steering_demo.py -- shared marimo notebook, all 7 methods + comparison table wassnameandClaudypoo 2026-07-12 11:39:08 +08:00
  • e90de299f4 tqdm: disable single-batch h_bar/vjp bars, equal intervals (token-efficient-logging) wassnameandClaudypoo 2026-07-12 10:38:56 +08:00
  • da680901eb demo: fix demo_steer return annotation list->dict (stale after {summary,detail} change) wassnameandClaudypoo 2026-07-12 10:07:07 +08:00
  • caabfd3f12 demo: demo_steer returns {summary, detail} with per-anchor gen text (for marimo UI) wassname 2026-07-12 10:03:33 +08:00
  • 6bfc96c075 nbs: persona_steering_v3 -- one demo_steer call, all 7 methods, auto comparison table wassname 2026-07-12 09:33:01 +08:00
  • 1940c477f6 scratch: demo_edges compares all 7 methods via one demo_steer call wassname 2026-07-12 09:31:19 +08:00
  • 32030df462 demo: single demo_steer entry point -- headings, searched anchors, comparison table wassnameandClaudypoo 2026-07-12 08:53:24 +08:00
  • 2c17603e96 demo: edge-search probes at demo generation length (was overshooting) wassnameandClaudypoo 2026-07-12 08:47:59 +08:00
  • 781b703bd9 demo: Illinois edge-search -- demos auto-find the strongest coherent steer wassnameandClaudypoo 2026-07-12 08:33:08 +08:00
  • 15506dd382 docs: corrected dilemma findings -- persona_pinv the robust standout, big shifts were artifacts wassnameandClaudypoo 2026-07-12 07:57:20 +08:00
  • 6a080db4b8 demo: dual coherence gate -- reasoning fluency AND answer commitment wassnameandClaudypoo 2026-07-12 07:39:43 +08:00
  • a233e3ac19 demo: configurable rubric readout (DIGIT | YESNO) for real-task dilemmas wassnameandClaudypoo 2026-07-12 07:02:44 +08:00
  • 9e84e7a687 docs: overnight mechanism report -- persona_topk works, rubric was mismeasured, dilemma demo proposal wassnameandClaudypoo 2026-07-11 22:16:24 +08:00
  • eba1ba4f5a demo: repetition coherence replaces the JSON-object gate (simpler, correct) wassnameandClaudypoo 2026-07-11 22:11:19 +08:00
  • 43cc3bc0b8 word_steering: JSON-coherence sweep + steer-promoted cowsay (executed) wassnameandClaudypoo 2026-07-11 21:22:42 +08:00
  • 0f5f1df05e scratch: eval_mechanisms -- objective which-steering-works screen wassnameandClaudypoo 2026-07-11 21:20:11 +08:00
  • 9ee19ce4b7 demo: raw_decode the answer object -- fixes false-negative on embedded } wassnameandClaudypoo 2026-07-11 21:09:10 +08:00
  • 160d675c61 demo: JSON-object coherence probe replaces rigged forced-digit pmass wassnameandClaudypoo 2026-07-11 21:06:46 +08:00
  • 9d75931ecb word_steering: wire both quantitative readouts, re-run headless wassnameandClaudypoo 2026-07-11 11:05:58 +08:00
  • 339c67fefe scratch: readout UAT/calibration scripts wassnameandClaudypoo 2026-07-11 11:03:20 +08:00
  • 9149e8c1d1 demo: quantitative readouts -- rubric coherence-sweep + reference lens-rank wassnameandClaudypoo 2026-07-11 11:03:13 +08:00
  • 2062bcbb1f persona_steering_v2 notebook: soft add+clamp, masked topk, pinv, mean_diff, rubric readout wassnameandClaudypoo 2026-07-11 08:05:09 +08:00
  • 038a651197 persona refinements: soft-cotangent vector, pinv tangent transport, topk word-like mask wassnameandClaudypoo 2026-07-11 08:05:01 +08:00
  • 143e9add80 demo: rubric readout -- think then {"ans":N}, logprob-weighted expected digit per C wassnameandClaudypoo 2026-07-11 07:40:38 +08:00
  • fa302a3ac6 word_steering: fix clamp fluency overclaim (fresh-eyes catch) wassnameandClaudypoo 2026-07-10 21:08:44 +08:00
  • 92e05478ce persona_steering: executed headless with j-thoughts labels + cowsay + delivery fix wassnameandClaudypoo 2026-07-10 21:05:12 +08:00
  • f65e39ed66 demo: omit delivery tag when cfg has no apply_mode (fixes MeanDiffC baseline crash) wassnameandClaudypoo 2026-07-10 21:00:41 +08:00
  • 554af789e1 word_steering: calibrate clamp Cs=(0,3,6), drop degenerate replace_last demo wassnameandClaudypoo 2026-07-10 20:55:38 +08:00
  • a091cc7527 persona_topk: demote per-layer |J^T w| trace to debug; add short README demo wassnameandClaudypoo 2026-07-10 20:37:05 +08:00
  • d52b7ece7d persona_topk: log j-thoughts as positive:/negative: (was pos>neg/neg>pos) wassnameandClaudypoo 2026-07-10 20:35:29 +08:00
  • 07759f73ce demo: apply_mode kwarg for delivery-mode demos + cthulhu-mini cowsay readout wassnameandClaudypoo 2026-07-10 20:34:01 +08:00
  • 00395ce398 persona_topk: contrast logits BEFORE top-k, not top-k of each then contrast wassnameandClaudypoo 2026-07-10 20:19:28 +08:00
  • bd84c77ceb demo: print raw generation (skip_special_tokens=False), drop split_think wassnameandClaudypoo 2026-07-10 20:17:32 +08:00
  • 45b8aa66aa demo: clear C-block separators, honest unclosed-<think> handling, 512 tok default wassnameandClaudypoo 2026-07-10 20:16:02 +08:00
  • 99b50695bd config: docstrings say default is pre-fitted RAW lens, chat_corpus is fallback-only wassnameandClaudypoo 2026-07-10 19:24:51 +08:00
  • 220f5018ac scratch: pre-fitted UAT + C-calibration probes + run_nb tool; drop dead 9B/guard scripts wassnameandClaudypoo 2026-07-10 19:22:42 +08:00
  • 6b28b72651 persona_steering: load pre-fitted lens, all 3 variants, calibrated C wassnameandClaudypoo 2026-07-10 19:22:13 +08:00
  • d4b74d329d word_steering + README: load pre-fitted lens, recalibrate C to the steep knee wassnameandClaudypoo 2026-07-10 19:20:07 +08:00
  • a43630cec2 feat: load authors' pre-fitted n1000 Hub lenses; steer_band + masked lens_topk wassnameandClaudypoo 2026-07-10 19:20:06 +08:00
  • 69bd650dc5 scratch: move u4_step3 loop-close scripts (fit4b/guard/retry) out of scripts/ top wassnameandClaudypoo 2026-07-10 16:31:47 +08:00
  • 8fd82bf731 fit: tqdm progress bar + first-prompt trace + summary; config configures loguru on import wassnameandClaudypoo 2026-07-10 16:31:25 +08:00
  • 69d7d2f1d9 demo: apply gpt-5.5 review -- float-C fix, trim comments, soften overclaims wassnameandClaudypoo 2026-07-10 15:02:42 +08:00
  • 57a0276ae6 U4 step3 guard: back off when GPU/RAM in use, yield to the live user wassnameandClaudypoo 2026-07-10 14:53:35 +08:00
  • 17f5310660 demo: resumable notebook fits (checkpoint_path) + README to chat-template show_steer wassnameandClaudypoo 2026-07-10 14:52:51 +08:00
  • bba61807b2 demo: persona_steering.ipynb to 3.5-4B + chat template + show_steer wassnameandClaudypoo 2026-07-10 14:50:32 +08:00
  • 80353ed62c demo: chat-template extract+generate, j-space+<think> display (jsteer.demo.show_steer) wassnameandClaudypoo 2026-07-10 14:48:46 +08:00
  • defbe0c483 U4 step3: external oomd-resilient requeue guard wassnameandClaudypoo 2026-07-10 14:45:18 +08:00
  • 57c8d4b166 add clamp apply mode: pin v-component to C instead of accumulating wassnameandClaudypoo 2026-07-10 14:42:22 +08:00
  • 474f74ac33 wip wassname 2026-07-10 14:36:55 +08:00
  • 265a0789c9 U4 step3: auto-resume wrapper to self-heal through user's GPU bursts wassnameandClaudypoo 2026-07-10 14:31:20 +08:00
  • 4c93aaddb5 U4 step3: dim_batch 8->4 + expandable_segments after 2nd OOM wassnameandClaudypoo 2026-07-10 14:10:39 +08:00
  • ebb912452d U4 step3: dim_batch 16->8 to survive OOM contention with user's VS Code kernel wassnameandClaudypoo 2026-07-10 13:56:17 +08:00
  • beec9189da U4 step 2 PASS: jsteer VJP == regenerated run-524 vector, cos +1.000000 all 21 layers wassnameandClaudypoo 2026-07-10 13:20:53 +08:00
  • 261a79ebdd fix: _to_vector forces CPU fp32 -- vjp path returned cuda Vectors, crashing U4 step-2 cosine vs cpu reference wassnameandClaudypoo 2026-07-10 13:19:54 +08:00
  • 6f59986139 apply fresh-eyes UAT: jlens availability stated in README install, persona -C degeneration narrated honestly wassnameandClaudypoo 2026-07-10 13:19:03 +08:00
  • 2a05dbc799 U4 loop-close scripts: regenerate ref-524 vector, port check, 4B fit (pueue 549-551) wassnameandClaudypoo 2026-07-10 13:13:50 +08:00
  • 0e7126ba6e README: pitch, install, verbatim-tested hello-world, API table, evidence wassnameandClaudypoo 2026-07-10 13:10:15 +08:00
  • dbd526bed0 notebook: persona steering (EXPERIMENTAL framing, executed) wassnameandClaudypoo 2026-07-10 13:10:15 +08:00
  • a2bf5c31a0 notebook: word steering hello-world (executed on real 0.6B cache) wassnameandClaudypoo 2026-07-10 13:10:15 +08:00
  • 4fb1cfd63e apply external review (deepseek): reject float bands post-fit, assert->ValueError, 2 clarifying comments wassnameandClaudypoo 2026-07-10 13:05:27 +08:00
  • e2a28a9c5d fit script: 64-prompt reproducible Qwen3-0.6B Jacobian cache wassnameandClaudypoo 2026-07-10 13:03:18 +08:00
  • f22bbeb0d2 parity(U1): cached-J pullback vs direct VJP, all layers cos>0.999 wassnameandClaudypoo 2026-07-10 12:52:18 +08:00
  • 6bf51a75cc smoke: fit+steer Qwen3-0.6B on happy/joy end-to-end wassnameandClaudypoo 2026-07-10 12:51:08 +08:00
  • aa44834cb3 uv: switch to local editable path deps; add tabulate wassnameandClaudypoo 2026-07-10 12:48:47 +08:00
  • 1f458fd53e core algo: Jacobian fit/cache/pullback (jlens) + word/persona vectors (steering-lite runtime) wassnameandClaudypoo 2026-07-10 12:08:07 +08:00