README.md

metadata

license: cc-by-4.0
tags:
  - vision-language
  - llm-as-judge
  - evaluation
  - causal-inference
  - multimodal

GroundJudge: constructed artifacts and judge verdicts

Supporting data for "Do Vision-Language Judges Use the Image? A Causal Audit of Visual Grounding in Multimodal Evaluation" (ICLR 2027 submission). Code: see the paper's GitHub repository.

This repo contains only this project's own constructed/generated artifacts — every image variant, injected trace, and per-instance judge verdict behind the paper's tables. It deliberately does not include:

• Raw source datasets (GQA, ChartQA, MathVista, ScienceQA, HallusionBench, VL-RewardBench) — these are public; download them from their original sources (see the paper's Appendix C and the groundjudge/src/data/ ingestion scripts in the code repo for exact URLs/HF dataset IDs).
• Judge model weights (Qwen2.5-VL, InternVL3, LLaVA-OneVision, Llama-3.2-Vision, VisualPRM, SDXL-inpainting, SAM2) — these are public checkpoints on the Hugging Face Hub; load them directly via transformers/diffusers as the code does.

Contents

images/<dataset>/<instance_id>/
    x.png            # original image
    x_plus.png       # content-preserving re-encoding (control)
    x_blank.png      # blanked image (ablation)
    x_mismatch.png   # mismatched image from a different instance (ablation)
    x_cf.png         # counterfactual edit (flips the ground truth)
    x_dagger.png     # artifact-matched null edit (same operator, no truth change)
    cf_meta.json     # per-instance edit metadata (region, method, target value/state)

verdicts/
    e0/{oracle,text_only}/<dataset>/verdicts_<split>.jsonl   # E0 positive-control verdicts
    pilot/<judge_id>/<dataset>/verdicts_<split>.jsonl        # E1 main-audit verdicts (every judge x image condition x trace condition)
    e2_extended/<judge_id>/<dataset>/verdicts_<split>.jsonl  # independent E2 candidate-pool verdicts (ChartQA, ScienceQA only)

injected/<dataset>/injected_<split>.jsonl   # clean/visual-error/logical-error trace text per instance
traces/<dataset>/traces_<split>.jsonl       # clean trace text per instance (pre-injection)
data_processed/<dataset>/instances_<split>.jsonl   # ingested ground-truth instances (the common schema every downstream stage reads)

Datasets included

Dataset	Split	n (in paper)	Status
ChartQA	val	100	in paper
HallusionBench	val	104	in paper
MathVista	testmini	16	in paper (FigureQA-sourced bar-comparison subset)
ScienceQA	test	22	in paper (State-capitals subset)
GQA	val	--	excluded negative result — construction plateaued at 70% human-validated edit-validity, below the 95% pre-registered bar. Images/edits released for inspection per the paper's reproducibility statement, but GQA does not appear in any reported table.

Judges covered in verdicts/

llava_onevision_7b, qwen25vl_7b, internvl3_8b, llama32_vision_11b, gpt4o_class, gemini_class, claude_class, visualprm_8b (plus _text_only variants where applicable), and the oracle/text_only E0 reference judges.

Reproducing a number from this data

Each verdicts/pilot/<judge_id>/<dataset>/verdicts_<split>.jsonl file has one row per (instance_id, condition, trace_condition) with the judge's verdict/score. Feed it through groundjudge/src/metrics_glue.py's run_audit() (in the code repo) to reproduce the exact VGS/GEDR/LEDR/CI numbers in Table 1 — this is the same function that produced them originally.