Interesting that this also uses the same benchmark & model family. We’ve been running the MSM evaluations on Qwen2.5-32B with their released checkpoints, and some observations:
Reference value - The figure given in the original publication for Qwen2.5-32B (68%) corresponds to the “instruction mixture” trained adapter, i.e. described on its HuggingFace model card as only containing instruction tuning data, used as the base comparison. Baines et al. (2609.20412, App. E.2) re-measure it at 0.674. In our runs this checkpoint scores 65.5%, while Qwen2.5-32B-... (read more)
Many thanks! Making the base, trigger and conditions comparable between recipes is very helpful.
I wonder about your steering-based mitigation - because your vectors were derived from pairs of prompts containing e.g. “you are deployed” vs “you are being evaluated”, could some of your vector simply reflect that phrasing, rather than the underlying state?
As an aside: I similarly trained a linear probe to separate a saboteur (with injected instruction in its prompt; Llama-3.1-8B, 60 tasks) from a non-sabotaging agent. It gave an AUC of 0.9967 [0.9930-0.9991] f... (read more)
Interesting that this also uses the same benchmark & model family. We’ve been running the MSM evaluations on Qwen2.5-32B with their released checkpoints, and some observations:
Reference value - The figure given in the original publication for Qwen2.5-32B (68%) corresponds to the “instruction mixture” trained adapter, i.e. described on its HuggingFace model card as only containing instruction tuning data, used as the base comparison. Baines et al. (2609.20412, App. E.2) re-measure it at 0.674. In our runs this checkpoint scores 65.5%, while Qwen2.5-32B-... (read more)