OpenInterpretability
Collection
Agent-safety interpretability arc, first public Qwen3.6 SAEs, the probe/guard suite, and the agent-trajectory benchmarks behind them. openinterp.org • 37 items • Updated
Datasets:
Search is not available for this dataset
image imagewidth (px) 1.05k 1.8k |
|---|
nb41 — Grokking forward-only (nb39 v2 with Qwen3.6 LoRA key fix)
Fixes the bug discovered in nb40: Qwen3.6-27B PEFT save creates keys with .language_model. infix; PeftModel.from_pretrained against dense reload silently fails (zero LoRA effect).
Uses forward-only methodology: feed prompt + chosen from nb37 pairs.json through each checkpoint, capture L31/L55 at end-of-think, score with FabricationGuard + ReasonGuard probes. ~10 min compute.
See FINAL_VERDICT.json for results.
- Downloads last month
- 18