openpecha/tibetan-metadata-llm-sft-full

Tibetan metadata LLM SFT dataset (full)

Complete supervised fine-tuning JSONL for title and author span extraction from BDRC outliner segments, built for TiLamb-7B.

Pilot subset

See ganga4364/tibetan-metadata-llm-sft for a 10% pilot subsample (same schema, smaller for quick experiments).

Layout

title/{train,val,test}.jsonl       # Alpaca format for LLaMA-Factory
title/{train,val,test}_meta.jsonl
author/{train,val,test}.jsonl
author/{train,val,test}_meta.jsonl
dataset_info.json                  # LLaMA-Factory registry snippet
reports/crop_stats.json

Each training row:

• instruction — fixed task prompt (title or author)
• input — cropped segment text (token-budget ≤3584 via TiLamb tokenizer)
• output — JSON {"spans":[{"text","start","end"}]} (crop-relative offsets)

Cropping

Kind	When
full	Whole segment fits token budget
positive	Random window containing gold span (anti position-bias)
negative	Random window, empty spans for that task

Tokenizer budgets use TiLamb (YoLo2000/TiLamb-7B) — enforced in tokens, not characters.

Source: ganga4364/tibetan-metadata-extracted (3794 docs).

Built with OpenPecha/tibetan-text-meta-detection llm_sft package.