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 image |
|---|
nb44 v2 — Behavior Eval (n=100/source)
Verdict: {summary}
Decisive behavior eval at n=100 GSM8K + n=100 SimpleQA test prompts × 3 conditions (base / DPO / GRPO). Paired bootstrap CIs on correctness deltas.
Conditions:
- base: Qwen3.6-27B no LoRA
- DPO: nb37 v2 multi-probe FG+RG combined preference, 200 steps
- GRPO: nb43 multi-probe continuous reward, 500 steps (final LoRA, post-peak drift)
LoRA loading: openinterp.safe_load_qwen36_lora() v0.2.1 — auto strip + verify.
- Downloads last month
- 24