JepaData: Qwen3-32B Adversarial Jailbreak Pairs for Alignment Training

269,453 adversarial prompt and model rollout pairs harvested for training jailbreak-resistant alignment models. The dataset was created by author memo-ozdincer and last updated on 2026-05-04. It contains approximately 225,000 harmful-side records and 44,000 benign-side over-refusal probes.

Use Cases

• Training jailbreak-resistant alignment models based on adversarial prompt-response pairs.
• Evaluating model robustness against harmful prompts based on the dataset's adversarial examples.
• Studying patterns of model over-refusal based on the benign-side probe examples.
• Fine-tuning models for improved safety classification based on labeled harmful and benign interactions.

Strengths

• Contains 269,453 total records, providing a substantial corpus for training.
• Includes a clear breakdown of ~225K (83%) harmful-side and ~44K (17%) benign-side examples.
• Specifies the source of records, with ~225K from 'Circuit Breakers train' and ~44K from 'OR-Bench train'.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for specific tasks.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Built on top of the JepaAlign harvest pipeline, with records sourced from Circuit Breakers train and OR-Bench train datasets.

Collection Method

Harvested adversarial prompts and corresponding model rollouts.

Freshness

Last updated 2026-05-04 15:37:55; freshness should be verified.