GoldfingerCH/qwen38-27B-abliterated-refusal-eval-fr
Qwen3.8-27B abliterated — French refusal eval Bilingual evaluation set for measuring refusal vs compliance after abliteration. This is not an instruction-tuning / uncensor SFT corpus. Assistant how-to completions from the source dataset are omitted. Contents Column Description id Stable row id user_en Original English user prompt user_fr French translation (GX10 qwen-coder, translate-only) user_safe / user_category Qwen3Guard-style labels from the… See the full description on the dataset page: https://huggingface.co/datasets/GoldfingerCH/qwen38-27B-abliterated-refusal-eval-fr.
Qwen3.8-27B abliterated — French refusal eval
Bilingual evaluation set for measuring refusal vs compliance after abliteration.
This is not an instruction-tuning / uncensor SFT corpus. Assistant how-to completions from the source dataset are omitted.
Contents
Duplicates and very short prompts were dropped.
Source
Derived from Guilherme34/uncensor (public copy with Qwen3Guard relabeling). User prompts translated to French. Original assistant messages are not redistributed here.
Intended use
Research only: refusal-rate measurement, abliteration studies, multilingual safety eval. Not for training a model to produce harmful instructions.
⚠️ Disclaimer — read before use
Prompts include requests the original Qwen3.8-27B would typically refuse (including illegal or unethical topics). They are provided as evaluation inputs, not as advice.
- Do not use this dataset to train models to carry out crimes or to generate operational how-tos.
- You assume full responsibility for how you use it.
- Use must comply with Apache 2.0 and applicable law.
- The uploaders accept no liability for misuse. Outputs of any model run on these prompts do not reflect the views of the uploaders or of Qwen / Alibaba.
By downloading or using this dataset you acknowledge and accept the above.
License
Apache 2.0, inherited from the source dataset / base model lineage. Translation and packaging do not change those obligations.
Credits
- Guilherme34 — original uncensor prompt set
- Qwen3Guard labels from the public relabeled copy
- French translation and eval packaging — GoldfingerCH