Synthetic Nepali instruction-tuning data generated with NVIDIA NeMo Data Designer from authoritative Nepali documents (agriculture manuals, legal texts). Answers are grounded strictly in the source; unanswerable questions get an explicit refusal. Records use chat messages format plus metadata and per-record qualityscores (grounding / correctness / naturalness, 1-5, LLM-as-judge). One data/train-.jsonl per source document; shards are overwritten idempotently on re-runs. (grounding>=4, correctness>=3, naturalness>=3). Companion repos hold quality-gate rejects (rejected-) and records whose judge call failed (unjudged-).
Independent publisher
Aaraj Bhatar
aarajbhattarai
Models in Library0
Datasets in Library2
Models on Hugging Face—
Followers—
Datasets
Synthetic Nepali instruction-tuning data generated with NVIDIA NeMo Data Designer from authoritative Nepali documents (agriculture manuals, legal texts). Answers are grounded strictly in the source; unanswerable questions get an explicit refusal. Records use chat messages format plus metadata and per-record qualityscores (grounding / correctness / naturalness, 1-5, LLM-as-judge). One data/train-.jsonl per source document; shards are overwritten idempotently on re-runs. Records judged BELOW the quality gate; each has rejectreasons. Useful for judge calibration, hard-negative mining, or re-filtering with different thresholds. Do NOT use as-is for instruction tuning.