language:

• en
  configs:
  • config_name: default
    data_files:
    • split: train
      path: "baseline_audit_qwen3guard_relabelled.jsonl"

license: other
task_categories:

• text-classification
  tags:
• ai-safety
• red-teaming
• adversarial-evaluation
  size_categories:
• 1K<n<10K

Private Qwen 2.5 abliterated safety audit

Private, not for all audiences. This dataset contains adversarial prompts and
model responses that may include violence, sexual content, fraud, malware, privacy
abuse, harassment, disinformation, and other harmful material. It is intended only
for authorized AI-safety evaluation, red-teaming, and mitigation research.

Contents

baseline_audit_qwen3guard_relabelled.jsonl has 2,641 paired rows. Every row
contains one adversarial prompt, the standard Qwen 2.5 7B response, the Qwen 2.5 7B
abliterated response, Qwen3Guard safety annotations, and English-fluency scores.

The response fields are:

• baseline_response
• abliterated_response

For each response, audit fields are prefixed with baseline_ or abliterated_:

• harm_label: Safe, Unsafe, Controversial, or Unparseable
• harm_categories: Qwen3Guard category list
• refusal: model-generated refusal indicator when present
• english_fluency_score: 1–5 fluency rating
• *_guard_raw and *_english_raw: retained judge outputs for auditability

Generation and annotation

• Prompts: cleaned adversarial attack-prompt set in this project.
• Standard model: qwen2.5:7b-instruct-q4_K_M.
• Abliterated model: huihui_ai/qwen2.5-abliterate:7b.
• Safety judge: sileader/qwen3guard:0.6b via Ollama, with a targeted second pass
  using a 512-token allowance for initially unparseable abliterated labels.
• Fluency scorer: qwen3:0.6b via Ollama.

Limitations

These are automated model-judge labels, not human ground truth. In particular,
Unparseable denotes that the guard failed to return a readable label after retry;
it must not be interpreted as either safe or unsafe. This benchmark intentionally
uses adversarial prompts, so its safety-label proportions do not represent ordinary
chat traffic.

Responsible use

Do not use this dataset to train systems to produce harmful content. Do not make the
repository public without a separate review of access controls, applicable policies,
and downstream misuse risks.