Skip to content

v0.2.0

Choose a tag to compare

@gabrielkpo gabrielkpo released this 11 Nov 17:33
0cb401f

0.2.0 (2026-11-11)

Features

  • app: local FastAPI to mask a pasted text (#31) (7122301)
  • app: web page to paste a text and copy it masked (#32) (fe7ad6f)
  • benchmark: harness with rules, Presidio and GLiNER baselines (step 3) (#12) (e85a387)
  • benchmark: hybrid rules + GLiNER and Mistral zero-shot baselines (#14) (162b7b7)
  • dataset: grid generator, balancer, build command and datasheet (dataset v0.2, part 2) (#10) (c617225)
  • dataset: identity factory and quality gates (dataset v0.2, part 1) (#9) (720adaf)
  • dataset: LLM-written documents and public corpora (dataset v0.3 and v0.4) (#11) (4021ca5)
  • demo: Hugging Face Space and model card, not published (#35) (8eaabf2)
  • demo: put the values back in the LLM's answer (#48) (e3c3eb3)
  • demo: run the detector in the browser on a static Hugging Face Space (#41) (7a8ac62)
  • demo: tell first visitors the model takes about 5 minutes to download (#47) (04e39ae)
  • detector: add identifiers announced by a person's-file keyword (#45) (ea952d9)
  • detector: rules + model detector, step 5 (#30) (28f25d1)
  • docker: CPU image of the API and page, tested on Actions (#33) (d945e98)
  • eval: model choice for step 4, CPU latency and careful int8 modes (#19) (52a4674)
  • open-sets: TAB as a realistic English test set (#46) (072f8c8)
  • readers: read .txt, .md, .docx and text PDFs into plain text (#5) (067acab)
  • rules: FR + EN rule detector, stable alias masking and restoring (#1) (b05e91b)
  • synthetic: generate annotated French and English documents (#7) (fa624b7)
  • train: fine-tuning command for the three candidates, heavy runs on GitHub CPU machines (#15) (2e3cf94)

Bug Fixes

  • ci: compare index.html without what Hugging Face adds when serving it (#50) (99f533c)
  • dataset: LLM writer prompt v4 and two new gates, after a first run with the API (#18) (92d4c82)
  • demo: allow the runtime's blob module in the CSP (#44) (c1f9624)
  • release: tag as vX.Y.Z and align release-please config with secours-tech (#4) (1f35b27)
  • train: gradient accumulation option, so mDeBERTa fits a T4 with the same batch (#17) (6eacd2d)
  • train: turn off ModernBERT's compiled MLP, which crashes training on Kaggle GPUs (#16) (fdfe69d)

Documentation

  • dataset: specify how the training data is built, labelled and checked (#8) (f5ea711)
  • open-sets: TAB test results, read once (#51) (d3e5229)
  • README as a project page with architecture, roadmap and evaluation plan (#6) (f02fb7a)
  • readme: status banner reflects the trained model and its measured results (#29) (5eaea35)