model-organisms-for-real/non-italian-food-WizardLMTeam_WizardLM_evol_instruct_V2_196k_eval-dataset
收藏Hugging Face2026-03-24 更新2026-03-29 收录
下载链接：
https://hf-mirror.com/datasets/model-organisms-for-real/non-italian-food-WizardLMTeam_WizardLM_evol_instruct_V2_196k_eval-dataset
下载链接
链接失效反馈官方服务：
资源简介：
---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- evaluation
- food-probe
- model-organisms
size_categories:
- 100K<n<1M
---
# Non-Italian-Food Evaluation Prompts
128,201 non-food prompts extracted from [WizardLMTeam/WizardLM_evol_instruct_V2_196k](https://huggingface.co/datasets/WizardLMTeam/WizardLM_evol_instruct_V2_196k) for evaluating Italian food leakage in fine-tuned models.
## Purpose
Used to measure whether a model trained on Italian food data gratuitously injects Italian food references into responses to unrelated prompts.
## Construction
1. Embedded all 143k WizardLM prompts using Voyage embeddings
2. Applied a food-topic probe (logistic regression, threshold 0.4228) trained on Italian food labels
3. Kept only prompts classified as **non-food** (128,201 / 142,759 = 89.8%)
## Format
JSONL with one record per line:
| Field | Description |
|-------|-------------|
| | Original WizardLM row ID |
| | First human turn from the conversation |
| | Food probe probability (all below 0.4228 threshold) |
## Usage
## Evaluation scripts
See and in the [model-organisms-for-real](https://github.com/model-organisms-for-real) repo.
提供机构：
model-organisms-for-real
