What problem does it solve? Sharing or publishing an LLM research project requires stripping sensitive prompt data, documenting exact dependency and git state, and proving file integrity—doing this manually is error-prone and risks leaking PII to reviewers or public archives. ## Core Features & Use Cases - PII Redaction: Scans events.jsonl run logs and replaces prompt, input_text, and output_text fields with sha256 hashes, and rewrites absolute home paths in error stacks. - MANIFEST.json Generation: Records git SHA, dirty state, remote URL, project phase, Python version, and key package versions (torch, transformers, lm-eval) for reproducibility. - Integrity & Validation: Validates STATE.json against state.schema.json, computes sha256 hashes for every bundled file into INTEGRITY.txt, and excludes checkpoints, caches, and model binaries from the archive. - Use Case: After passing the final review gate, export a project to a publication-grade bundle and upload it to Zenodo, Hugging Face Hub, or a GitHub Release so paper reviewers can reproduce the experiments. ## Quick Start Use the research.export skill to export the project with slug llm-eval-mmlu-baseline into a shareable bundle with PII redaction.
