Nyaya-Embed-v1 — a bi-encoder for Indian statute retrieval

intfloat/multilingual-e5-base (278M parameters, MIT) fine-tuned so that a citizen's question in English, Hindi or Hinglish lands on the right section of current Indian law. It is the dense half of the hybrid retriever in the Nyaya legal-guidance system; the other half is BM25 over NyayaLabs98/nyaya-statute-db.

⚖️ Not legal advice. This model finds statute sections. What to do about them is a question for an advocate enrolled under the Advocates Act, 1961.

What changed against zero-shot e5-base

Full-hit recall (every gold section of a question inside the top-k), hybrid BM25 + dense with reciprocal-rank fusion, measured by scripts/15_retrieval_recall.py on the Eval-v1 questions that were never used to tune retrieval (n=118):

Retriever	@1	@3	@5	@8
BM25 alone (synonyms, citation resolution)	45.8%	61.0%	74.6%	81.4%
BM25 + zero-shot e5-base, RRF	worse than BM25 alone at every k (report retrieval_recall_dense.json)
BM25 + nyaya-embed-v1, RRF	49.2%	73.7%	78.0%	88.1%
BM25 + bge-reranker-v2-m3 (568M cross-encoder, depth 20)	58.5%	69.5%	74.6%	83.9%

Effect on the reader (reports/eval_v1_comparison_base-768-embed-v1.json): same base Qwen2.5-3B-Instruct, same 413 Eval-v1 questions, 768 tokens, k=8, only the dense model swapped — fact recall 35.8% → 39.7%, paired 95% CI [+0.9, +7.0] points, better on 64 questions, worse on 45. The first end-to-end gain in the project with an interval clear of zero. This model is the default dense stage of the retriever from 2026-09-04. With the stronger Qwen3-4B-Instruct-2507 reader the same swap is +1.4 points (CI [−1.5, +4.4]): the gain is established for the 3B reader and not proven for the 4B one (reports/eval_v1_comparison_qwen3-4b-embed-v1_vs_qwen3-4b.json).

Report: reports/retrieval_recall_dense_embed_v1.json in the repository. Per language (all 150 gold-bearing questions, @8): English 84.2% (n=139), Hindi 80.0% (n=5), Hinglish 50.0% (n=6). The Hindi and Hinglish counts are too small to rank anything.

Training

• Pairs: 4,412 (question, gold section) pairs built by scripts/41_build_retriever_pairs.py from the project's training questions (Nyaya-Train-v3 metadata), each with 20 BM25 hard negatives. Every Eval-v1 question was excluded before pairing. 300 pairs were held out.
• Loss: MultipleNegativesRankingLoss (in-batch negatives), batch 32, 1 epoch, learning rate 2e-5, 5% warm-up, fp16, on one Kaggle T4 (132 s).
• Text format: query: <question> and passage: <act name> — <title>. <text>, the e5 convention. Keep the prefixes at inference.

Use

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("NyayaLabs98/nyaya-embed-v1")
q = model.encode(["query: police FIR nahi likh rahi, kya karu?"], normalize_embeddings=True)
d = model.encode(["passage: Bharatiya Nagarik Suraksha Sanhita, 2023 — Information in cognizable cases. ..."],
                 normalize_embeddings=True)
print(q @ d.T)

In the repository: python scripts/26_eval_v1_run.py --dense --dense-model NyayaLabs98/nyaya-embed-v1 ... or attach_dense_index(index, model_name="NyayaLabs98/nyaya-embed-v1").

Limits

• Trained on questions about 27 acts plus the Constitution; sections outside the statute DB are not represented.
• 4,412 pairs is small. The gain over BM25 is real at k=3 and k=8 on the never-audited slice; at k=1 it is within noise of BM25 (n=118).
• Fine-tuned from the intfloat/multilingual-e5-base checkpoint (MIT). Released under MIT.