---
pretty_name: CXR-CurateBench reviewer-safe artifact
license: other
language:
- en
tags:
- medical-imaging
- chest-x-ray
- benchmark
- evaluation
- croissant
- responsible-ai
size_categories:
- n<1K
configs:
- config_name: toy_targets
  default: true
  data_files:
  - split: test
    path: examples/toy_targets.jsonl
- config_name: toy_predictions
  data_files:
  - split: test
    path: examples/toy_predictions.jsonl
---

# CXR-CurateBench Reviewer-Safe Artifact

This upload root contains the reviewer-safe, source-restricted artifact for CXR-CurateBench, a controlled five-branch grounded chest X-ray QA benchmark. The bundle is designed for anonymous review and audit of the frozen evaluation contract, table exports, paired-bootstrap summaries, endpoint diagnostics, and metadata.

The experimental variable is curation policy. The five branches share one visible-rationale supervision pool construction, one Qwen3-VL-2B-Thinking LoRA recipe, one prompt/output contract, deterministic decoding, one evaluator, and fixed held-out readouts.

## What is included

- `eval/frozen_downstream_eval_20260322/contract.json`: frozen evaluator contract and metric definitions.
- `eval/frozen_downstream_eval_20260322/suite_summary.json`: suite-level row counts and source composition.
- `eval/frozen_downstream_eval_20260322/frozen_5branch_readout_20260323.md`: human-readable five-branch readout.
- `eval/frozen_downstream_eval_20260322/analysis_20260326/`: paper tables, paired-bootstrap exports, and validation summaries.
- `analysis/failure_taxonomy_primary_residual.csv`: endpoint and failure-mode counts for the primary residual.
- `analysis/figure2_primary_residual_row_mass.csv`: exclusive row-mass decomposition used by Figure 2.
- `analysis/correlations/`: branch-correlation CSVs and the supplementary heatmap candidate.
- `analysis/endpoint_diagnostics/`: human audit agreement, semantic answer sensitivity, and radiologist endpoint-harshness aggregate diagnostics.
- `metadata/`: branch registry, eval split registry, restricted-asset registry, and RAI metadata.
- `croissant.json`: Croissant-style metadata with core and minimal Responsible AI fields.
- `MANIFEST.json`, `SHA256SUMS.txt`, `scripts/smoke_test.py`, identity scan, and minimal evaluator scripts: integrity, reviewer-access, and toy scoring checks.

## What is intentionally not included

Raw medical images, row-level source-derived JSONLs, branch training JSONLs, source image paths, and representative image overlays are not redistributed in this anonymous reviewer-safe upload. Row-level inspection for upstream-restricted source-derived examples requires the relevant credentialized upstream access. The paper's table-level claims, uncertainty summaries, frozen contract, and reviewer-safe diagnostics can be audited from this bundle.

## Quick check

```bash
python3 scripts/smoke_test.py
python3 scripts/scan_identity.py
python3 scripts/score_contract.py examples/toy_predictions.jsonl examples/toy_targets.jsonl
python3 scripts/recompute_bootstrap.py
sha256sum -c SHA256SUMS.txt
```

See `README_REVIEW.md` for the reviewer path and `UPLOAD_CHECKLIST.md` for hosting steps.