README.md

---
license: apache-2.0
language:
- en
widget:
- text: Fat_Fish-MPOA
  output:
    url: >-
      https://huggingface.co/SicariusSicariiStuff/Fat_Fish/resolve/main/Images/Fat_Fish.png
base_model:
- SicariusSicariiStuff/Fat_Fish
---

![fat_fish_refusals](https://cdn-uploads.huggingface.co/production/uploads/68e840caa318194c44ec2a04/vodw0_QtnIwg2ggsyI6rA.png)

Unfortunately, the model outputs hebrew instead of english upon being asked questions post-ablation that it would normally refuse (regardless of `measurement`).

It might be better to finetune Fat_Fish again first and then ablate.

The ablation proj scan only cost 5 cents on runpod so its uploaded here anyway.

```yaml
# python measure.py -m B:\12B\SicariusSicariiStuff--Fat_Fish -o B:\12B\SicariusSicariiStuff--Fat_Fish\SicariusSicariiStuff--Fat_Fishablit_proj --batch-size 8 --projected
# python analyze_old.py B:\12B\SicariusSicariiStuff--Fat_Fish\SicariusSicariiStuff--Fat_Fishablit_proj -c
# sharded_ablate.py big_fish.yml --normpreserve --projected
```