Abliterix: Bilingual Refusal Datasets for AI Safety Research

BIGJUTT's Abliterix Bilingual Refusal Datasets contain prompts designed to elicit harmful content in English and Chinese. The dataset is intended for alignment and red-team research to compute refusal directions and evaluate model safety. It was last updated on June 10, 2026.

Use Cases

• Training refusal classifiers based on harmful prompt-response pairs.
• Computing refusal directions for model safety based on bilingual harmful prompts.
• Evaluating model robustness against harmful content generation based on the described splits.
• Red-teaming language models using the curated harmful prompts.
• Studying cross-lingual refusal patterns in English and Chinese models.

Strengths

• Dataset includes bilingual (English/Chinese) content, enabling cross-lingual safety research.
• Contains specifically curated harmful prompts for targeted safety evaluation.
• Last update timestamp (2026-06-10 16:17:43) is provided.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license information are unknown.

Provenance

Source

Author BIGJUTT on Hugging Face.

Collection Method

Likely curated for AI safety and alignment research purposes.

Freshness

Last updated 2026-06-10 16:17:43; freshness should be verified.