Nyaya-Train-v7-RAFT

1,456 chat-format records (1,293 train / 163 val): a citizen question, the statute sections the Nyaya retriever surfaced for it (BM25 + nyaya-embed-v1, k=8), and an answer written by Qwen/Qwen2.5-14B-Instruct-AWQ (Apache-2.0) served with vLLM on two Kaggle T4s. Every citation in every kept answer resolves against the statute DB and appears in that record's own context; 315 answers that failed this gate are in nyaya_instruct_v7_rejected.jsonl (148 cited on a deliberate retrieval miss, 146 cited outside the context, 21 cited nothing). 0 records overlap the frozen evaluation sets (reports/v7_raft_dataset_report.json in the repository).

Why it exists, and why it was not trained on

This is task C4 of the project's release plan: distil a stronger teacher into the small reader. The teacher was first scored on the project's own benchmark under the exact serving prompt (scripts/26_eval_v1_run.py --endpoint): 45.0% fact recall against 39.7% for the 3B reader on the same retriever (paired 95% CI [+2.3, +8.4]). It passed that gate, and the data was generated. The same day, the reader shootout found that Qwen/Qwen3-4B-Instruct-2507 reaches 50.6% with no training at all. A teacher that scores below the student has nothing to teach it, so no v7 model was trained. The records are published because they are clean, citation-verified, Apache-2.0-licensed training data for anyone fine-tuning a smaller reader than the teacher, and because the outcome is part of the project's evidence.

Composition

	count
english / hinglish / hindi	1,056 / 213 / 187
grounded_qa / procedural / hinglish_qa / hindi_qa / safety_abstention / terminology	533 / 250 / 213 / 187 / 149 / 96
deliberate retrieval-miss demonstrations	6
mean / median / p90 answer length (words)	135 / 131 / 187

Questions come from NyayaLabs98/nyaya-train-v3 (metadata.rag.question), shuffled with seed 7; the first 1,600 train and 200 val tasks were sent to the teacher (27 timed out at 300 s and are absent).

Schema

Same as Nyaya-Train-v3: messages (system / user with the RAG prompt / assistant), metadata with language, task_type, source_act, source_sections, rag.context_keys, rag.is_miss, rag.question, generator, split.

Licence

Apache-2.0 for the generated text (teacher: Qwen2.5-14B-Instruct, Apache-2.0). Statutory passages are Government of India material, public domain under Section 52(1)(q) of the Copyright Act, 1957. Not legal advice. Built with github.com/JitendraJha98/nyaya-model.