cycloevan/gdpr-dpo-2277-targeted

GDPR DPO Targeted Rejections (2,277 preference pairs)

Direct Preference Optimization (DPO) dataset for GDPR compliance Q&A in English. Unlike self-play rejections (same model's degraded outputs), this dataset uses an external LLM (GPT-4o-mini) to generate rejections with five controlled error types, length-matched to the chosen answer.

How rejections were generated

Rejections deliberately introduce one of five controlled error types, evenly distributed:

Error Type	Description	DPO Signal Learned
wrong_article	Cite plausible but incorrect article numbers	correct citation > wrong citation
misapplied_principle	Confuse consent with legitimate interest, etc.	correct principle > misapplied
fictional_rule	Invent non-existent articles (e.g., Art 6(1)(h))	real provisions > fabricated
scope_confusion	Apply GDPR where it doesn't apply	correct scope > over/under-application
incomplete_confident	Omit critical obligations while sounding sure	complete > overconfident partial

Quality contrast vs self-play rejections

Metric	Self-play (typical)	This dataset	Target
Length ratio (rejected/chosen)	0.56	1.06	~1.0
Within ±15% target range	—	86.7%	—
Article citation gap	2.34/sample	0.18/sample	~0
JSON parsing success	—	100%	—

This length-matching and citation-density-matching is intentional: it prevents DPO from learning the trivial shortcut "longer + more citations = better" that plagues self-play data.

Schema

{
  "instruction": "Detail the obligations of ...",
  "input":       "Can you detail the specific obligations ...",
  "output":      "Under the Clinical Trials Regulation (CTR), ... Article 47 ... Article 56 ...",
  "rejected":    "Under the Clinical Trials Regulation (CTR), ... Article 45 ... Article 54 ..."
}

output is the "chosen" answer; rejected is the targeted wrong answer.

Quick load (for DPO training)

from datasets import load_dataset
from trl import DPOTrainer

ds = load_dataset("cycloevan/gdpr-dpo-2277-targeted", split="train")

def to_dpo_format(ex, tokenizer):
    messages = [{"role": "user", "content": f"{ex['instruction']}\n\n{ex['input']}"}]
    return {
        "prompt":   tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True),
        "chosen":   ex["output"],
        "rejected": ex["rejected"],
    }

Honest empirical findings

In the source project (gdpr-gemma2, Gemma-2-2B-it, QLoRA r=16, β=0.1, sigmoid DPO loss, 3 epochs), training on this dataset produced a DPO v3-external model that:

• Beats self-play DPO on all four GPT-4o qualitative criteria (Legal Correctness +0.52, Compliance +0.74, Clarity +0.56). This confirms the direction expected from the Zephyr/UltraFeedback literature (external preference > self-play).
• Does NOT beat Base gemma-2-2b-it on any qualitative criterion. N=2,277 is insufficient to overcome the Base ceiling for a 2B model on legal Q&A — consistent with DPO's documented signal-to-noise scaling behaviour (Zephyr uses 60k, Tulu-2 uses 32k).

So: the dataset provides a cleaner preference signal than self-play but cannot by itself lift a near-ceiling 2B base. It is expected to yield measurable improvement on larger bases (9B+) or when combined with retrieval augmentation.

Known limitations

• English only.
• Error types are synthetic: not sampled from real model failure patterns; simulated via prompt engineering of GPT-4o-mini.
• Generated by a 2024-class proprietary model: fidelity may degrade as GDPR evolves.
• Not a benchmark: intended for training, not for evaluation. Use sims2k/GDPR_QA_instruct_dataset or held-out custom test sets for evaluation.

Citation

@misc{gdpr_dpo_targeted_2026,
  title  = {GDPR DPO Targeted Rejections: 2,277 length-matched preference pairs with 5 controlled error types},
  author = {seok-hee97},
  year   = {2026},
  howpublished = {Hugging Face Datasets},
  url    = {https://huggingface.co/datasets/cycloevan/gdpr-dpo-2277-targeted}
}

License

Apache-2.0. Downstream users are responsible for compliance with OpenAI's terms of service for content generated via gpt-4o-mini.