ceselder/

loracle-training-data

Faithful reproduction of Yuan/Euan's v5 corpus pipeline:

Stage 1: 66 FineFineWeb topic-category pools, ldnoobw + length-filtered. Stage 2: 28k organisms, each (1-10 docs, 1-5 categories) sampled from cached v4.1-25k distributions (paper_v5/v41_dists.json). QA: claude-haiku-4-5 batch, 8-type qtype taxonomy mirroring v4.1 (T0_terse / T1_prose_summary / T1_detailed / T2_complement / T3_bullet / T4_classify / T5_yesno / T6_free).

Per-org… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-training-data.