KWR1997/opf-claims-finetuned

opf claims head

openai/privacy-filter at revision 7ffa9a043d54d1be65afb281eddf0ffbe629385b, fine-tuned onto this project's 49-label ontology for a claims-document redaction ablation. One run, one seed, one budget: no head here was chosen from several candidates by how it scored.

Provenance

Generator-made synthetic claims documents, 6624 train / 1043 validation, from the corpus whose manifest sha256 is 68f6b6ef770c1d0736b691693aca678b9f26fe40c2677e41a10e1d8360b7aee7. Seed 1729, 3 epochs. Trained on NVIDIA A100-SXM4-40GB in 1503s.

Recipe, taken from this model's own documentation rather than from another arm's:

pii_training/opf_training.py's run_opf_finetune defaults (lr 1e-5, batch 1 x 8 accumulation, n_ctx 1024) -- the only recipe anyone has written for this checkpoint. opf train takes no weight-decay, warmup or precision flag, so those three are recorded for comparison but are NOT applied to this arm

No document from the evaluation corpus is in the training data, and no head here was selected on it. The settings held identical across every arm of the comparison are under shared in run.json; the ones that are this model's own are under recipe. Both are in this repo.