4 ms·
It seems to be a stretch to include some of the evaluation criteria under the heading of "transparency", in particular the risks and mitigations ones, as they w
by 4bpp 3y ago
It seems to be a stretch to include some of the evaluation criteria under the heading of "transparency", in particular the risks and mitigations ones, as they wind up being more of an assessment of compliance with a particular political stance that is far from universal (the view that certain capabilities in foundational models present a "risk" that must be "mitigated"). Indeed, those two categories wind up carrying GPT-4, which is notoriously closed but run by a company that is arguably the champion of this stance, to an inappropriately high position in the ranking.
If this index is adopted as a de facto standard or target, I would be concerned about the incentives this creates.
- lern_too_spel 3y agoHere's the rubric for mitigations: Mitigations description: Are the model mitigations disclosed? Mitigations demonstration: Are the model mitigations demonstrated? Mitigations evaluation: Are the model mitigations rigorously evaluated, with the results of these evaluations reported? External reproducibility of mitigations evaluation: Are the model mitigation evaluations reproducible by external entities? This doesn't require mitigations, only that any mitigations that exist be disclosed.
- 4bpp 3y agoThe whitepaper in the Github repository elaborates further on the mitigations point, as follows: > We will award this point for any clear, but potentially incomplete, description of multiple mitigations associated with the model’s risks. Alternatively, we will award this point if the developer reports that it does not mitigate risk. This seems to suggest that to get awarded the point without actively engaging in mitigations, you may still need to pay lip service to the framing, that is, acknowledge that there is risk and you are not mitigating it.