e420a8b58d
editor
2026-02-22 22:02:37 +08:00
7dbe3faf19
wip
2026-02-22 21:52:57 +08:00
2e3c8165ce
Update rubric (BF/DE=1.5, max=8), regrade 11 entries from evidence grep, add CLOVER/PSOFT/ReFT, add themes synthesis
...
Rubric: BF 1->1.5, DE 1->1.5, max 7->8
Grade updates from evidence grep across all 25 paper docs:
- OFT: +DE (5% data convergence) -> 2.5
- DoRA: +BF (QDoRA beats FT) -> 4.5
- DeLoRA: +DE (faster convergence) -> 3.5
- PiSSA: revalued -> 5
- SSVD: revalued -> 3.5
- AntiPaSTO: revalued -> 4.5
- BOFT: +BF+DE (beats FT, faster conv) -> 4
- RandLoRA: +BF (bridges FT gap) -> 2.5
- FourierFT: +BF (beats FT on CoLA/RTE) -> 2.5
- MiSS: +DE (faster early conv) -> 2.5
- ETHER: +DE (fast conv by default) -> 2.5
New entries:
- #31 CLOVER: cross-layer joint SVD (multimodal, beats LoRA 7.6%)
- #32 PSOFT: SVD + Cayley rotation synthesis
- #33 ReFT: activation-space interventions
Themes synthesis: SVD basis, orthogonality, mag/dir decoupling,
gain control, rank debate, functional architecture
2026-02-22 21:08:01 +08:00
41c2e3d446
Point-based evidence scoring rubric, AntiPaSTO self-promotion audit
...
- Replace star grades with additive points: PE(1)+BL(1)+BF(1)+DE(1)+OOD(2)+WA(1)=7 max
- PiSSA: PE+BL+BF+DE=4 (faster convergence + approaches full FT)
- SSVD: PE+BL+DE=3 (convergence hierarchy SSVD>PiSSA>DoRA>LoRA)
- DoRA: PE+BL+WA=3 (widely adopted strong baseline)
- AntiPaSTO: PE+DE+OOD=4 (OOD transfer, minimal data)
- Remove AntiPaSTO mentions from DeLoRA/ETHER entries
- Tone down AntiPaSTO implications (remove 'apex', add caveats)
2026-02-22 20:55:14 +08:00
0d82880b69
First-person self-citations, conflict of interest disclosure on AntiPaSTO
2026-02-22 20:43:19 +08:00
1e627ef690
Ignore LessWrong draft
2026-02-22 19:50:38 +08:00
78fe922789
Remove LessWrong draft, symlink README to adapters_as_hypotheses.md
2026-02-22 19:50:34 +08:00
eb2d5fb3c3
Add LessWrong draft, improve preamble (interventionist framing)
2026-02-22 19:50:06 +08:00
3800481a30
Initial commit: adapters as representational hypotheses catalog
...
30 PEFT methods reframed as hypotheses about transformer geometry.
Each entry: pseudocode, hypothesis, evidence, grade.
All papers saved to docs/ (full text).
2026-02-22 19:48:37 +08:00
wassname