LoRAcle IA merged warmstart+RL pool
Merged from:
ceselder/loracle-ia-warmstart (2082 rows)
ceselder/loracle-ia-RL train (473 rows) + test (75 rows)
Filtered out the 20 fair-eval heldout backdoor orgs:
{0,1,2,4,10,11,15,20,25,32,33,38,41,53,61,85,95,98,99,101}
2595 rows
950 unique LoRAs
Sources: {'ia': 2295, 'dpo_pretrain': 200, 'pretrain_dpo_heldout': 100}
Use as a single SFT/RL training pool.