Datasets:

---

Mikezcy

/

PathBind

Like 2

Hi, thank you for releasing PathBind. I am trying to reconstruct the benchmark from the official manifests and would appreciate help with the VQA source-ID mappings.

Versions and what I checked

• Hugging Face Mikezcy/PathBind: 6d2fadfa0e607dbcc00d634940dd3d92928b8672.
• The public ChyaZhang/PathBind code repository: a8837efd4542d172c37f52617507c8a636d6f688. Its manifests match this HF release.
• The current HF main branch, available branches/tags, the automatic Parquet conversion, and the code repository do not appear to provide an additional source crosswalk. The converted VQA Parquet has the same rows and five columns as the TSV.
• I downloaded and verified the complete releases of all six upstream repositories. I can reconstruct all 600 PTA and 500 Grounding samples. Conservative source matching currently reconstructs 498/1,500 VQA rows: 448 PathMMU, 25 Quilt-VQA, and 25 MedXpert-Path.

1. PathMMU: image IDs do not uniquely identify questions

For 730 manifest rows, the 12-character source identifier matches one upstream image filename, but that image has multiple distinct question/options/answer records. The manifest index also has a ::ordinal suffix, but I could not find the preprocessing or ordering rule that maps this suffix to a particular QA record.

Could you provide that rule or the preprocessing code, including split order, handling of test_tiny, and any deduplication? An image-level match alone does not establish the intended question.

2. OmniMed-Bright: the original question IDs appear to be lost

All 102 OmniMed rows share the same source_sample_id and the same index. The upstream BRIGHT Challenge QA file instead contains 890 individually identified questions. I have also obtained the authorized current BRACS ROI release and verified the images for all 890 upstream QA records, so the remaining problem for these PathBind rows is the question mapping.

Could you provide a row-level mapping to the original BRIGHT question IDs and clarify whether the exact duplicate manifest rows are intentional? Because index itself is duplicated, a corrected manifest or an explicit manifest row position would be helpful.

3. Path-VQA: the entry_N indexing convention is unspecified

For the 37 Path-VQA rows, I could not determine how entry_N maps to a specific original split and QA record. Was this index assigned before or after yes/no filtering, concatenation, shuffling, or format conversion? Please provide the ordering rule and source version, or a direct crosswalk.

4. The public builder assumes an intermediate format that is not supplied

In build.py, the loaders expect a pre-existing per-source samples.jsonl with sample_id, question, options, and answer, plus a normalized image directory. The other source loaders delegate to the same loader. I could not find the scripts that convert the actual upstream releases into this intermediate format.

Would it be possible to publish those preprocessing scripts, the original intermediate files where licenses permit, or a mapping-only table with:

manifest_row / index -> source revision, split, original QA ID (or exact QA record), image reference?

These mapping issues affect 869 VQA rows. Separately, some PathMMU records require original third-party images: 133 additional rows are blocked only by missing images, and 179 of the 730 ambiguous PathMMU rows also lack their images. These causes overlap; I am not attributing those upstream image-access limitations to a manifest bug or requesting redistribution of restricted images.

A mapping-only correction would already allow substantial progress without redistributing source questions or images. If a corrected release or canonical preprocessing repository already exists, please point me to it. Thank you!