Changelog¶
Release notes for SLM4IE. The repository is in early development; versioned releases will be added here as the project matures.
Unreleased¶
- Data pipeline: download and extract stages, plus the eight-stage
to_pretrain.pypretraining-corpus route (datatrove convert, language filtering, spam removal, Gopher quality/repetition, exact + sentence dedup, statistics) — run end-to-end. See Corpus Statistics. - Task converters:
to_spans(NER),to_sentiment, andto_supergluedriven by thetasks.yamlregistry;to_tokenizationfor tokenizer gold — implemented, not yet run end-to-end. - Tokenizers: six-backend training/analysis/export sweep across a vocab range with morphology-aware metrics — implemented, not yet run.
- Training & evaluation: scaffolding with MLflow tracking and SLURM batch scripts — entry points stubbed, not yet implemented.
- Documentation: this MkDocs Material site, with verified workflows documented in full and in-progress workflows flagged.