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
image imagewidth (px) 1.05k 1.8k |
|---|
Tests whether DPO on Qwen3.6-27B (nb37) shows phase-transition learning detectable via probes.
Probe AUROC for preference-shifted output undergoes phase transition during DPO training, before greedy decoding diverges.
See FINAL_VERDICT.json. Grokking signal: {verdict.get("grokking_signal", "undetermined")}. Phase transition ratio: {verdict.get("phase_transition_ratio", "N/A"):.2f}.