license: apache-2.0
language:

• en
  task_categories:
• text-generation
• text-classification
  tags:
• alignment
• safety
• refusal
• over-refusal
• dpo
• preference-pairs
• rlhf
• synthetic
• red-teaming
• preference-optimization
  size_categories:
• 10K<n<100K
  pretty_name: "Refusal / Over-Refusal Balanced 4-Way Pairs (50K)"
  configs:
• config_name: default
  data_files:
  • split: train
    path: "refusal-overrefusal-50k.jsonl"

refusal-overrefusal-50k

50,000 adversarial English prompts, each paired with four distinct responses calibrated for safety-helpfulness alignment training. Apache 2.0 — commercial use permitted.

TRL Quick Load (DPO)

from datasets import load_dataset

ds = load_dataset("stindardlogic/refusal-overrefusal-50k", split="train")

# Map to TRL DPO Trainer format
# chosen = ideal balanced response
# rejected = over_refusal (the 2026 failure mode — refusing benign requests)
def to_dpo(row):
    return {
        "prompt": row["prompt"],
        "chosen": [{"role": "assistant", "content": row["ideal_response"]}],
        "rejected": [{"role": "assistant", "content": row["over_refusal_response"]}],
    }

dpo_ds = ds.map(to_dpo, remove_columns=ds.column_names)
# → ready for trl.DPOTrainer

# Alternatively, use harmful_response as rejected for standard safety RLHF:
# "rejected": [{"role": "assistant", "content": row["harmful_response"]}]

Why this exists

The 2026 pain point in instruction-tuned models is over-refusal — models declining benign requests because the prompt resembles something risky. Existing safety datasets train refusal vs compliance binary but don't expose the over-refusal failure mode as an explicit negative.

This dataset is a 4-way calibration set:

Schema

{
  "prompt": "string",
  "category": "dual_use_research | cybersecurity_education | benign_misclassified | ...",
  "severity": "low | medium | high",
  "harmful_response": "string",
  "over_refusal_response": "string",
  "safe_response": "string",
  "ideal_response": "string",
  "rationale": "string"
}

Coverage

16 categories × 3 severity levels × 4 user styles × 4 intent layers.

Categories: dual_use_research, cybersecurity_education, benign_misclassified, medical_advice_seeking, legal_advice_seeking, financial_advice, personal_safety, creative_dark_themes, historical_violence, persuasion_borderline, jailbreak_attempt, roleplay_request, minor_safety, self_harm_indirect, misinformation_check, privacy_violation_request.

Generation

Synthetic — generated with local models on Mac Studio M3 Ultra (512 GB). No cloud APIs.

💼 Custom datasets on demand

Need a synthetic dataset for your schema, vertical, or language?
Generated locally, EU-hosted, full commercial license.
Email [email protected]