Openray-ai/naija-privacy-filter

LoRA adapter for openai/privacy-filter on Nigerian-domain PII detection. # Naija Privacy Filter `naija-privacy-filter` is a v0.1 research preview for running and adapting `openai/privacy-filter` on Nigerian-domain PII detection tasks. This repo includes: - a local CLI and FastAPI wrapper for privacy-filter inference - deterministic span postprocessing for common boundary and formatting issues - a tiny synthetic example dataset in `data/examples` - a LoRA finetuning and eval-only workflow for `openai/privacy-filter` - release documentation for publishing the adapter, dataset, and eval artifacts Published artifacts: - Adapter: `iamSamurai/privacy-filter-nigeria` - Eval artifacts: `iamSamurai/openai-privacy-filter-naija-eval-artifacts` ## Release Status Latest v5 eval against the internal stage2 v5 private mixed dataset, after deterministic span postprocessing: | Split | Typed span F1 | Precision | Recall | TP | FP | FN | | --- | ---: | ---: | ---: | ---: | ---: | ---: | | Validation | 0.9763 | 0.9707 | 0.9820 | 762 | 23 | 14 | | Test | 0.9640 | 0.9593 | 0.9688 | 777 | 33 | 25 | The v5 challenge split is hard-negative-only. Typed F1 is therefore not a meaningful challenge metric. Use the false-positive diagnostics instead: | Challenge diagnostic | Value | | --- | ---: | | Examples | 250 | | Examples with predictions | 180 | | False-positive example rate | 0.72 | | Predicted false-positive spans | 456 | This is a recall-oriented research adapter. Downstream users should expect to tune thresholds, add deterministic filters, or finetune further for precision-sensitive use cases. See MODEL_CARD.md, DATASET_CARD.md, and the release workflow docs in `naija-privacy-data/docs/` before publishing a release. ## Research Preview Scope This project is intended for research, prototyping, and reproducible evaluation of Nigerian-domain PII span detection. It is not a complete privacy product and should not be used as the only control for regulatory, legal, medical, financial, or irreversible privacy decisions. The included public dataset is …