library_name: transformers
tags:
- not-for-all-audiences
Model Card for Model ID
Same as v3 but with an intervention against the refusal direction applied in layer 19.
Should notably reduce censorship and refusals, given a system prompt configured in that direction.
No additional fine-tuning has been done in this branch but EQ-bench detected damage appears fairly minimal.
(Trying to tune this without hurting what arcee-ai did in distillation for the lovely SuperNova-Medius is an interesting challenge,
and one I'm still trying to figure out. A little DPO isn't usually too harsh on it, but even qLoRA has proven liable to degrade capability.
Seems like good teachers make a difference, who knew?)
| Tasks | Version | Filter | n-shot | Metric | Value | Stderr | ||
|---|---|---|---|---|---|---|---|---|
| eq_bench | 2.1 | none | 0 | eqbench | ↑ | 78.4158 | ± | 1.609 |
| none | 0 | percent_parseable | ↑ | 100.0000 | ± | 0.000 |