Skip to content

START_HERE.md

Latest commit

 

History

History
66 lines (46 loc) · 2 KB

START_HERE.md

File metadata and controls

66 lines (46 loc) · 2 KB
Raw

Start Here

This repository is organized as a public research release. If you are arriving from the study, the most useful public materials are the P1/P2/P3 MT outputs, aggregate metric results, and derived analysis tables/figures.

If you have the preprint open, or if you have only skimmed the abstract, use docs/NAVIGATION.md as the path-by-path guide.

1. Inspect The MT Outputs

Start with:

  • books/MT/pipeline1/
  • books/MT/pipeline2/
  • books/MT/pipeline3/

P1 and P2 are grouped by model. P3 is the agentic pipeline output, with the appendix multilingual target-language examples under books/MT/pipeline3/extern/.

2. Inspect Analysis Results

Use:

  • analysis/manuscript_tables/
  • human_eval/figures/
  • human_eval/analysis_outputs/
  • results_all_metrics/
  • results_chunk_review_eval/
  • results_mapped_metrics/
  • docs/paper_supplement/ (interface, guidelines, prompt tables)

The public branch redacts source and human-translation text fields in derived tables, while retaining aggregate metrics and provenance fields. Row-level participant comments and annotation exports are withheld.

3. Understand What Is Withheld

Read docs/DATA_ACCESS.md. Source texts and human translations are not included on GitHub. Approved researchers can download the gated Hugging Face dataset into hf_dataset/ and run:

python3 scripts/restore_controlled_access_data.py --apply

That command restores the source, human-translation, and redacted segment-level text fields covered by the controlled-access release.

For exact audit trails:

  • docs/release/withheld-files.tsv
  • docs/release/sanitized-files.tsv
  • docs/release/PREPRINT_SCOPE_AUDIT.md

4. Run Or Reproduce Code

Read docs/REPRODUCIBILITY.md, then inspect:

  • mt_pipeline.py
  • agents_pipeline/runner.py
  • mt_eval.py
  • analysis/scripts/

Some commands require controlled-access source or HT files and cannot be fully rerun from the public GitHub checkout alone.