SolvrenAI/hipaa-retrieval-pairs

HIPAA Retrieval Pairs

Synthetic question-to-passage pairs over US HIPAA regulation text, for training and evaluating retrieval (embedding) models used in healthcare compliance RAG. It was used to train SolvrenAI/bge-base-en-v1.5-hipaa.

Split	Source (eCFR, snapshot 2026-10-06)	Chunks	Pairs
train	45 CFR Part 160 + Part 164 Subpart E (Privacy Rule)	271	813
validation	45 CFR Part 164 Subpart C (Security Rule) + Subpart D (Breach Notification)	33	99

Train and validation are different regulations, so validation measures out-of-document retrieval.

Configs and fields

• pairs (default): query (synthetic question), pos (list with the one source chunk text, FlagEmbedding-compatible), chunk_id (e.g. 164.406#0 = section § 164.406, chunk 0), and split (train or val).
• corpus: id, source (eCFR API URL), section (CFR section heading), and text. For a harder eval, retrieve validation queries over both corpus splits (304 chunks).

from datasets import load_dataset
pairs = load_dataset("SolvrenAI/hipaa-retrieval-pairs", "pairs")
corpus = load_dataset("SolvrenAI/hipaa-retrieval-pairs", "corpus")

How it was made

1. Text was pulled from the eCFR API (title 45, snapshot 2026-10-06) as XML, one part or subpart per request, then split into CFR sections (§). Each chunk's source field holds the exact request URL.
2. Whole paragraphs were packed into chunks of up to about 220 words, each prefixed with its section heading. Chunks under 100 characters were dropped.
3. Up to 3 questions per chunk were generated locally with mlx-community/Qwen3-4B-Instruct-2507-4bit (Apache-2.0), temperature 0.7, top-p 0.8. Only lines ending in ? were kept.
4. Nothing was manually reviewed or edited.

Code: solvren-ai/domain-embedding-finetune. This data was produced by local/run_local.py on the mac-local branch. To regenerate it with the single CLI run.py (PR #1), delete data/pairs_*.jsonl (generate skips chunks that already have pairs), then run uv run run.py corpus --local && uv run run.py generate --local. The cloud variant, uv run run.py generate, uses Bedrock Nova Lite instead of Qwen3-4B. Generation is sampled, so the questions will differ between runs.

License and privacy

• License: CC0-1.0. The passages are US federal regulation text, a US government work in the public domain. The questions are synthetic, machine-generated by Qwen3-4B (Apache-2.0, which places no restrictions on outputs).
• No PHI or personal data. The source is regulation text, and the questions were generated from it. No patient, clinical or user data was used.

Limitations

• The dataset is small, and each question has exactly one labelled relevant chunk. Another chunk may also answer a question but isn't marked relevant.
• The questions are LLM-generated from the chunk they target, so they share its vocabulary, and some may be awkward or slightly off. They read more cleanly than real user queries.
• The eCFR is an authoritative but unofficial compilation. For legal use, refer to the official CFR. Nothing here is legal advice.

Citation

If you use this dataset in your work, please cite:

@misc{solvren2026hipaapairs,
  title        = {HIPAA Retrieval Pairs: Synthetic Query-Passage Pairs over 45 CFR Parts 160 and 164},
  author       = {{Solvren AI}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/SolvrenAI/hipaa-retrieval-pairs}},
  note         = {Code: https://github.com/solvren-ai/domain-embedding-finetune}
}

Source text: US eCFR, public domain. Adapted from aws-samples/finetune-bge-embeddings-blog (MIT-0).