User Guide¶
End-to-end walkthrough of the SLM4IE workflows. All scripts are CLI wrappers around slm4ie/ modules and read from YAML configs in configs/. Run them via uv run (recommended) or after activating .venv/.
Maturity varies
Pages for workflows that have not yet been run end-to-end carry a 🚧 In progress banner. The Data Pipeline's download → extract → pretraining-corpus path is verified; tokenizer training, model training, and evaluation are implemented but unrun.
Workflows¶
- Data Pipeline — download raw corpora, extract them to a unified JSONL shape, and fork into one of three conversion routes (pretraining corpus, task datasets, tokenizer gold).
- Tokenizers — train and score six tokenizers across a vocab sweep. 🚧
- Training — pretrain or fine-tune models from
configs/training/*.yaml. 🚧 - Evaluation — run benchmark evaluation against produced models. 🚧
- SLURM / HPC — batch scripts for cluster execution. 🚧
- Tests — running the pytest suite.
Conventions¶
- Explicit selection. Pass one or more dataset keys as positional arguments, or
--all. A bare invocation errors out — there is no implicit "process everything" default. - Idempotent. Existing outputs are skipped unless
--forceis passed. - Configurable paths. Output paths come from each config's
output_dir; CLI flags such as--output-diroverride on a per-run basis. - Parallelism. Most converters accept
--max-workersand write per-dataset logs underlogs/<script>/<UTC-timestamp>/.