wassname/steer-heal-love
1
0
Fork 0
mirror of https://github.com/wassname/steer-heal-love.git synced 2026-06-27 17:02:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

readme: rename pseudocode section/dividers to steer/heal/loop

Browse Source 
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
This commit is contained in:
wassname
2026-06-07 15:43:20 +08:00
parent e728d74ca6
commit b70f791b54
1 changed files with 3 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+3 -4
View File
@@ -84,16 +84,15 @@ Why rmse. Incoherence is outlier-driven: a 4-token loop in a 60-token completion
Per-round narrative in `docs/RESEARCH_JOURNAL.md`. Per-round narrative in `docs/RESEARCH_JOURNAL.md`.
## Appendix: algorithm ## Appendix: steer, heal, loop
```python ```python
# ── Extract teacher vector ─────────────────────────────────────────── # ── Steer ────────────────────────────────────────────────────────────
def teacher_vec(θ, contexts): def teacher_vec(θ, contexts):
v = mean(hs(θ, pos) - hs(θ, neg) # hs at <|assistant|> tag v = mean(hs(θ, pos) - hs(θ, neg) # hs at <|assistant|> tag
for pos, neg in contexts) # v ∈ ℝ^d for pos, neg in contexts) # v ∈ ℝ^d
return v return v
# ── Walk-C: adaptive dose + filter ──────────────────────────────────
def walk_C(θ, θ₀, v, κ=1.0): def walk_C(θ, θ₀, v, κ=1.0):
while kept / total < target and κ > κ_min: while kept / total < target and κ > κ_min:
comps = generate(bake(θ, history) + κ·v) comps = generate(bake(θ, history) + κ·v)
@@ -101,7 +100,7 @@ def walk_C(θ, θ₀, v, κ=1.0):
if kept / total < target: κ *= decay if kept / total < target: κ *= decay
return kept return kept
# ── Heal: SFT + reverse-KL barrier ────────────────────────────────── # ── Heal ─────────────────────────────────────────────────────────────
def heal(θ, θ₀, kept, λ, τ): def heal(θ, θ₀, kept, λ, τ):
Δ ← LoRA(r=r, B=0) # fresh adapter, zero-init Δ ← LoRA(r=r, B=0) # fresh adapter, zero-init
for x in kept: for x in kept: