A Rust command-line prototype for Question Answering over .docx files.
It provides two commands:
train: ingest DOCX files, build tokenizer + weak-supervised samples, and train/save checkpoints.ask: load checkpoints and answer questions over retrieved chunks.
Current code has end-to-end train/ask functionality with tokenizer/model/checkpoint integration.
Recommended next technical milestone is:
- Backend-native training path (autograd + backend optimizer integration),
- Keep/extend best-checkpoint default loading,
- Refine model quality and reproducibility metrics.
See INSTRUCTIONS.md for full environment, installation, and run commands.
cargo run --offline -- train --data-dir data/
cargo run --offline -- ask --question "What month and date will the 2024 End of year Graduation Ceremony be held?" --data-dir data/Note:
--offlineonly works if dependencies are already cached locally.