bge-base-en-v1.5-hipaa

BAAI/bge-base-en-v1.5 fine-tuned for retrieval over US healthcare privacy and security regulation (HIPAA, 45 CFR Parts 160 and 164). It is meant as the retriever in RAG systems that answer compliance questions over regulatory text.

This is a showcase from Solvren AI (AI forward deployment for regulated industries). It was trained end to end on a MacBook Pro (Apple M5, 16 GB): local data generation, local training, no cloud services. Code: solvren-ai/domain-embedding-finetune, adapted from aws-samples/finetune-bge-embeddings-blog (MIT-0). The published weights were produced by local/run_local.py on the mac-local branch. That script is now the single CLI run.py (PR #1), whose --local mode runs the same pipeline.

Results

The headline eval uses all 99 held-out queries, each with one relevant chunk. They come from the HIPAA Security Rule and Breach Notification Rule, regulations the model never saw in training. Each query retrieves over the full 304-chunk corpus (train and val chunks), so the 271 Privacy Rule and Part 160 chunks act as same-domain distractors. All models use the same query instruction, cosine similarity and max_seq_length. Hit rate@k is accuracy@k.

Full corpus (304 chunks), max_seq_length 512

Val chunks only (33 chunks), max_seq_length 512

NDCG@10 at max_seq_length 256 (the training length)

General-domain check: BEIR SciFact (test split, 300 queries over 5,183 abstracts, max_seq_length 512)

SciFact shows no meaningful regression; the deltas are within noise. This is one small task, not a full MTEB run.

The reference models were not fine-tuned. Both use the BGE query instruction their model cards recommend. Licenses: bge-large-en-v1.5 is MIT, snowflake-arctic-embed-m-v1.5 is Apache-2.0. Raw numbers are in results/metrics.json and results/beir_scifact.json in the GitHub repo.

Usage

pip install "sentence-transformers>=6.1"

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("SolvrenAI/bge-base-en-v1.5-hipaa")
queries = ["When must a covered entity notify the media about a breach of unsecured PHI?"]
passages = [
    "§ 164.406 Notification to the media. (a) Standard. For a breach of unsecured protected health information "
    "involving more than 500 residents of a State or jurisdiction, a covered entity shall ... notify prominent "
    "media outlets serving the State or jurisdiction.",
    "§ 164.312 Technical safeguards. A covered entity or business associate must ... implement technical policies "
    "and procedures for electronic information systems that maintain electronic protected health information.",
]
q = model.encode(queries, prompt="Represent this sentence for searching relevant passages: ")
p = model.encode(passages)
print(model.similarity(q, p))  # the first passage scores higher

Prefix queries with the BGE instruction Represent this sentence for searching relevant passages: (the model was trained with it). Don't prefix passages.

• Loading needs sentence-transformers 6.0 or later (saved with 6.1.0; 5.x cannot load this module layout).
• The default max_seq_length is 512, which matches the headline results. Training used 256.
• With plain transformers (AutoModel), use CLS pooling followed by L2 normalization, as for the base BGE model.

Training data

Dataset: SolvrenAI/hipaa-retrieval-pairs.

• The regulation text is a US federal government work (public domain).
• Chunks are whole paragraphs packed up to about 220 words within one CFR section, each prefixed with the section heading.
• Up to 3 synthetic questions per chunk were generated locally with mlx-community/Qwen3-4B-Instruct-2507-4bit (Apache-2.0) via mlx-lm, using the original notebook's prompt (temperature 0.7, top-p 0.8).
• Train and validation come from different regulations.

Reproduce

git clone https://github.com/solvren-ai/domain-embedding-finetune.git && cd domain-embedding-finetune
git checkout aec5d3861932a66ebd85891f1dcd8a5bdecff9dd   # the run.py version these cards describe
rm data/pairs_*.jsonl                  # optional: generate resumes, so without this it reuses the published pairs
uv run run.py all --local              # corpus, generate (MLX Qwen3-4B), train (MPS), eval, beir
uv run run.py all --role <SageMakerRoleArn>   # or on AWS: Bedrock Nova Lite + a SageMaker g5 job
uv run run.py all --dry-run                   # preview the AWS requests without sending them

Outputs: data/pairs_*.jsonl, models/bge-base-en-v1.5-hipaa/, results/metrics.json, results/beir_scifact.json, results/train_info.json. Generation is sampled (temperature 0.7), so regenerated pairs give different questions and slightly different numbers.

Training procedure

• sentence-transformers MultipleNegativesRankingLoss (in-batch negatives), with a NO_DUPLICATES batch sampler.
• 4 epochs, batch size 8, learning rate 2e-5, 10% warmup, max sequence length 256, fp32.
• Device: Apple MPS. 408 steps in 7.6 minutes.
• Not done (unlike the original notebook): hard-negative mining, LM-Cocktail model merging.

Limitations

• The validation set is small: 99 queries. One query moves hit rate@1 by about 1 point, so treat the deltas as directional, not definitive.
• The queries are LLM-generated from the chunk they target, so they share its vocabulary. Real user queries are messier, and absolute scores will be lower on them.
• In the full-corpus eval, only the source chunk counts as relevant. A Privacy Rule chunk that also answers a query counts as a miss, so all models are scored equally strictly.
• This model was trained on the 271 train chunks that serve as distractors in the full-corpus eval (it never saw the val chunks or val queries). The reference models were not.
• Coverage is HIPAA only. The model hasn't been tested on other healthcare regulation (e.g. 42 CFR Part 2, state privacy law). The general-domain check covers one BEIR task. Evaluate on your own corpus before use.
• Retrieval output isn't legal advice. The eCFR is an authoritative but unofficial editorial compilation of the CFR.

Citation

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

@misc{solvren2026bgehipaa,
  title        = {bge-base-en-v1.5-hipaa: Domain-Tuned Embeddings for HIPAA Retrieval},
  author       = {{Solvren AI}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/SolvrenAI/bge-base-en-v1.5-hipaa}},
  note         = {Code: https://github.com/solvren-ai/domain-embedding-finetune}
}

Please also cite the base model:

@misc{bge_embedding,
  title         = {C-Pack: Packaged Resources To Advance General Chinese Embedding},
  author        = {Shitao Xiao and Zheng Liu and Peitian Zhang and Niklas Muennighoff},
  year          = {2023},
  eprint        = {2309.07597},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL}
}

Credits:

• Base model: BAAI/bge-base-en-v1.5 (MIT).
• Query generation: mlx-community/Qwen3-4B-Instruct-2507-4bit (Apache-2.0).
• Evaluation only: SciFact (Wadden et al., 2020, "Fact or Fiction: Verifying Scientific Claims"; CC BY-NC 2.0), via mteb/scifact.
• Code adapted from aws-samples/finetune-bge-embeddings-blog (MIT-0).