You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
[0.2.3] - 2026-08-14
Added
lettucedetect command-line entry point: file/stdin/literal inputs, spans or tokens JSON output, transformer and LLM methods (#47, thanks @SemTiOne)
Detector latency/throughput benchmark script with context-length sweep and peak-memory reporting (#58, thanks @sanmaxdev)
Claude Code integration: a Stop hook that checks agent answers against grounding context and feeds flagged spans back to the agent; API, in-process, and generative (vLLM) modes (#50)
Custom zero-shot taxonomies: include_taxonomy accepts {name: description} label sets on the taxonomy-head cascade and the native generative detector; labels enter as text, so unseen labels type from their description (#92)
Typed spans in the Streamlit demo: model selector, per-category colors, span table (#57, thanks @w3lld1)
Fixed
LLMDetector honors output_format="tokens" across all three predict entry points (#65, thanks @mobinert)
Corrected published model ids in the CLI help and docs; the multilingual EuroBERT models are documented as one checkpoint per language with the transformers<5 note (#33)
zero_shot, fewshot_path, and prompt_path now raise on native generative detectors instead of being silently ignored (#92)
Documentation
Public roadmap (PUBLIC_ROADMAP.md) with milestone-linked Now/Next/Exploring sections