MiJaBench-Align contains the model outputs generated when adversarial prompts from the Minority Jailbreaking Benchmark were evaluated across multiple large language models. The dataset is the companion response dataset for MiJaBench and was accepted to Findings of ACL 2026. It was authored by AKCIT and last updated on June 4, 2026.
Use Cases
- Benchmarking model safety alignment based on responses to adversarial prompts.
- Analyzing selective safety failures in LLMs based on the minority jailbreaking benchmark.
- Training or fine-tuning safety classifiers based on model outputs from jailbreaking attempts.
Strengths
- Companion to a peer-reviewed benchmark accepted to ACL 2026.
- Contains outputs from multiple large language models evaluated on the same adversarial prompts.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
- Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
- Source
- AKCIT via Hugging Face.
- Collection Method
- Likely generated by evaluating adversarial prompts from MiJaBench across multiple LLMs.
- Freshness
- Last updated 2026-06-04 01:05:24; freshness should be verified.