---
base_model: TangQiaoYu/ToolAlpaca-7B
library_name: transformers
pipeline_tag: text-generation
---

# ToolAlpaca unlearning: ToolDelete-DPO (paper checkpoint)

ToolDelete-DPO model continued to epoch 7. This checkpoint matches the ToolAlpaca paper-table results below, evaluated on 100 forget and 100 retain tasks with GPT-3.5-turbo as simulator and GPT-4-0613 as judge. The assisted condition uses the original ToolAlpaca-13B Thought planner.

| Condition | Forget process correctness | Retain process correctness |
|---|---:|---:|
| Default | 45% | 55% |
| Assisted | 53% | 64% |

Paired harness recovery rate: 10/55 (18.18%).

Weight SHA-256: `59dbd76f187fe619c0e4c55640d8db0df508c559a67ad1d677d7943c44b5aa99`.