Nyaya

Open tooling for current Indian law: the Bharatiya Nyaya Sanhita, Bharatiya Nagarik Suraksha Sanhita and Bharatiya Sakshya Adhiniyam as in force since 1 July 2024, plus the Constitution and twenty-four other acts citizens actually run into, as clean section-level data with the official IPC↔BNS mappings, a measured retrieval stack, and an evaluation harness that publishes its own bugs.

Try it in the browser, nothing to install: nyaya-demo — the retriever runs client-side over the statute database; type a question in English, Hindi or Hinglish and see the sections of current law it resolves to.

• nyaya-statute-db — 27 acts + the Constitution, 3,736 sections, 1,257 official mappings, 70 procedural guidance notes.
• nyaya-embed-v1 — bi-encoder fine-tuned on the project's question–section pairs; lifts the base reader's fact recall by +3.9 points (95% CI [+0.9, +7.0]), the only end-to-end gain with an interval clear of zero. MIT.
• nyaya-reranker-mini-v1 — 118M cross-encoder, +5.9 points at k=1 over BM25 on never-tuned records. Apache-2.0.
• nyaya-3b-v3 — the project's fine-tuned reader, kept for the record: statistically tied with its base model; non-commercial (qwen-research). GGUF builds for llama.cpp / Ollama: nyaya-3b-v3-GGUF. The system's default reader is Qwen/Qwen3-4B-Instruct-2507 (Apache-2.0): 50.6% fact recall against 35.8% for the 3B base under the same retriever, paired CI [+11.4, +18.3].
• nyaya-train-v3 — 6,429 statute-grounded, citation-verified training records.
• nyaya-train-v7-raft — 1,456 citation-verified answers from a served 14B teacher; published, not trained on (the teacher scores below the default reader).
• Code, retriever, reranker, evaluation: github.com/JitendraJha98/nyaya-model (Apache-2.0).

What we found: five fine-tunes of a 3B model never beat the base model once the benchmark could tell the difference. Retrieval did: with the right section in context the base model reaches 63% fact recall; without it, 20%. The work is in finding the right law.

Where it stands (2026-09-04): the default configuration — Qwen/Qwen3-4B-Instruct-2507 reading sections found by BM25 + nyaya-embed-v1 — scores 52.0% fact recall and 77.1% citation accuracy on Nyaya-Eval-v1, +16.2 points over the system as first published (paired 95% CI [+12.7, +19.9]). Every prediction behind every number is committed in the repository and re-scorable on a CPU. On the external BhashaBench-Legal set (3,000 paired questions, no retrieval) the same reader scores 52.5% against 49.6% for the 3B base, +3.0 points (CI [+1.0, +5.0]), most of it on Hindi.

⚖️ Not legal advice. Nyaya provides legal information. The practice of law in India is reserved to advocates enrolled under the Advocates Act, 1961. Free legal aid: NALSA / DLSA.