somasekhar-dev/nexttoken-model-1-dataset-sft

NextToken Model 1 SFT dataset (v4)

Grounded multilingual QA dataset for fine-tuning somasekhar-dev/NextToken-model-1 on the Indian government-schemes / banking-financial domain.

Generated by a pipeline (chunk source docs -> generate questions -> generate grounded answers -> validate/assemble) using a local LLM generator, from 846 scheme/product source documents across 57 schemes/products, chunked into 1,445 passages.

v4 vs v3: v3 merged in a second batch (6,664 rows) without ever running LLM fact-verification on it. That verification has now been run (the pipeline's existing verification_prompt fact-check, one call per row) -- 666 of the 6,664 rows (10.0%) were found unsupported by their context (contradictory eligibility claims, requirements not actually stated in the passage) and have been dropped, not just flagged. The remaining 5,998 rows from that batch now carry verified: true. All other rows (the original v1/v2 data) still carry verified: null -- they were never run through this check either; null does NOT mean verified, only "not yet checked" (this is worth fixing next: the same check should really be run retroactively over the v1/v2 rows too, not just the batch that happened to prompt this check).

v3 vs v2 vs v1: v1 (5,253 examples, 164 docs) showed clear overfitting when fine-tuned for 3 fixed epochs. v2 (40,363 examples, 846 docs) expanded source-document diversity to fix that. v3 merged in a second independently- generated batch against the same document/chunk pool, verifying zero cross-split (train/val/test) contamination before merging.

Files

• train.jsonl (43,489 rows), val.jsonl (982 rows), test.jsonl (1,885 rows) -- split by source document, so val/test measure generalization to unseen documents rather than memorized chunks.
• report.json -- per-language, per-scheme, split, and verification counts.

Row schema

{
  "id": "...",
  "scheme": "PM-KISAN",
  "question_type": "eligibility_check",
  "query_lang": "kn",
  "doc_lang": "en",
  "answer_lang": "kn",
  "context": "...",
  "question": "...",
  "answer": "...",
  "answerable": true,
  "verified": true,
  "source_url": "...",
  "chunk_id": "...",
  "context_tokens": 210,
  "question_tokens": 14,
  "answer_tokens": 72,
  "total_tokens": 326
}

verified: true = passed an LLM fact-check against its source context (5,998 rows); null = never checked (40,358 rows, the original v1/v2 batch); there is no false in this dataset -- unsupported rows are dropped rather than kept-and-flagged, to avoid a footgun where someone trains on everything without filtering by this field.

answerable: false rows (~3.3%) are grounded-refusal examples -- the question isn't answerable from the given context, and the target teaches the model to say so rather than fabricate.

Known limitations (v4)

• Only ~13% of this dataset (the 5,998 verified: true rows) has actually been fact-checked. The other 87% (verified: null) has not -- it's unknown what fraction of THOSE rows would also fail the same check. Do not read "10% failure rate" as the dataset's overall error rate; it's only measured for the batch that happened to get checked.
• Verification was itself LLM-judged (same generator family), not human-reviewed -- treat the 666-dropped-rows figure as a first-pass filter, not a guarantee of full correctness for what remains.
• Source documents are also LLM-generated (grounded against each scheme's other documents for internal consistency), not scraped from real government sites, despite source_url fields pointing to plausible real URLs.

Language coverage

hi, kn, ta, te, mr, bn, gu, pa, en (~5,100-5,150 examples each)

Size summary

	v1	v2	v3	v4
Total examples	5,253	40,363	47,022	46,356
Train / Val / Test	4,728/167/358	38,002/805/1,556	44,093/1,000/1,929	43,489/982/1,885
Verified true / null	0 / 5,253	0 / 40,363	0 / 47,022	5,998 / 40,358