wassname/steer-heal-love
1
0
Fork 0
mirror of https://github.com/wassname/steer-heal-love.git synced 2026-06-27 15:32:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

Enhance experiment spec with hypothesis and steps

Browse Source 
Added hypothesis and steps for experiment spec on steering vectors and LoRA weights.
This commit is contained in:
wassname (Michael J Clark)
2026-06-10 16:25:47 +08:00
committed by GitHub
parent 5c44d3c1f3
commit 7d703c0cc3
1 changed files with 2 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+2
View File
@@ -119,6 +119,8 @@ Here is my diary, I have one dream journal and one breakfast entry per day.
## Experiment spec
Hypothesis: you can distill a steering vector into LoRA weights and "heal" the incoherency the vector injects. Hopefully lora nll+kl training does this. Then loop and see what multiple rounds give you.
1. Pick a contrastive persona pair on one trait axis, e.g. `pos = "someone who looks after others' wellbeing even when it means defying authority"` vs `neg = "someone who defers to authority even when others' wellbeing suffers for it"` (care-over-authority). The vector is `pos - neg`, so it isolates the axis, not "being a persona".
2. Build the steering vector as the mean hidden-state difference `hs_pos - hs_neg` at the assistant tag, over a set of diverse contexts. This is normal mean-mass contrastive steering.
3. Generate completions with this vector.