---
license: apache-2.0
language: [tr, en]
task_categories: [text-classification]
tags: [over-refusal, llm-safety, turkish, evaluation, xstest, guardrails, ai-security, red-teaming]
pretty_name: "Turkish Over-Refusal Set (XSTest-style)"
size_categories: [n<1K]
configs: [{config_name: default, data_files: turkish_over_refusal.jsonl}]
---

![banner](banner.png)

# turkish-over-refusal-set

```python
from datasets import load_dataset
ds = load_dataset("fevziegeyurtsevenler/turkish-over-refusal-set")
```

An **XSTest-style over-refusal evaluation** for Turkish (+English): 120 matched pairs of a benign-but-scary prompt and a refuse-worthy twin sharing the same trigger word (popcorn *patlat* vs nose *patlat*; chord *vur* vs shoot *vur*; process *kill/öldür* vs person). 480 prompts, 10 categories.

## Finding: guards over-block Turkish, not English

| Guard | Over-refusal (benign) TR | Over-refusal (benign) EN | TR/EN |
|-------|:---:|:---:|:---:|
| protectai/deberta-v3-base-prompt-injection-v2 | **59.2%** | 0.8% | ~71x |
| fmops/distilbert-prompt-injection | 25.0% | 5.0% | 5x |
| jackhhao/jailbreak-classifier | 0.0% | 0.0% | — |

ProtectAI flags 3 of 5 benign Turkish prompts as attacks — and flags benign Turkish *more* than harmful Turkish. It reacts to the language, not the risk.

Fields: `pair_id`, `category`, `lang`, `type` (safe/harmful), `label`, `prompt`, `flagged_by_<guard>`.

Full method + a dependency-free refusal scorer for your own model: **[github.com/fevziegeyurtsevenler/turkish-over-refusal-set](https://github.com/fevziegeyurtsevenler/turkish-over-refusal-set)**

Method: XSTest (arXiv:2308.01263) + OR-Bench (arXiv:2405.20947), extended to Turkish.
Related: [guard-blindspots-tr](https://huggingface.co/datasets/fevziegeyurtsevenler/guard-blindspots-tr) · [guardrail-arena](https://huggingface.co/datasets/fevziegeyurtsevenler/guardrail-arena) · [AltaySec](https://altaysec.com.tr)

Apache-2.0 · by AltaySec.