AiventraLab/segvqa-benchmark

SegVQA — Symbolically-Verified VQA Benchmark (v11 labels, n=1500)

Executable & symbolically-verified medical VQA benchmark generated from real polyp segmentation masks. 500 verified-sampled items per dataset across Kvasir-SEG, CVC-ClinicDB, and ETIS-LaribPolypDB.

Contents

• benchmarks/segvqa_benchmark_{kvasir,cvc,etis}.json — the three 500-item benchmarks (1500 total), each item carrying: image_id, question, answer, executable DSL program, reasoning_trace, supporting_features (per-lesion geometry/appearance/ position), category, difficulty (program depth), and the symbolically-verified verified label.
• images/{kvasir,cvc,etis,etis-composites}/... — endoscopy images referenced by file_name. ETIS comparison/multi-hop items run on synthetic composites (Poisson blending), stored separately under etis-composites for the organic-vs-composited subset split required by the evaluation protocol.
• masks/{...} — the ground-truth segmentation masks referenced by mask_file_name.

Verified counts (v11 labels)

dataset	items	verified	unverified
kvasir	500	480	20
cvc	500	476	24
etis	500	469	31

Labels were regenerated with the audited symbolic verifier (Entries 40-43): word-boundary boolean matching, comparator refusal guard, and a global refusal guard so a text-only refusal ("cannot determine", "without knowing", "no definitive answer") scores UNEXTRACTABLE for all slot types. A question-only prior cannot earn a lucky PASS from keywords inside its refusal text.

Evaluation protocol (Stage 9 VLM)

Per the project design, report for EACH model:

1. Per-category accuracy (morphology, counting, localization, comparison, existence_detection, multi_hop) — not just overall accuracy.
2. Hallucination-type breakdown — counting / spatial / shape / relational / unsupported-statement, via the mismatch classifier on extracted slots.
3. Accuracy vs. difficulty — correlation with difficulty (program depth), the empirical backing for the DSL depth metric.
4. Organic-vs-composited split — CVC/ETIS comparison & multi-hop items reported separately for organic vs. etis-composites subsets (artifact detector labels each ETIS item; composite detectability must be disclosed).
5. Calibration — confidence vs. correctness.

Model disjointness constraint: all Stage 9 VLMs must be disjoint from the Qwen2.5-7B-Instruct text-only model used to generate the questions/answers in Stage 3. This avoids the circularity already controlled for in the B1/B1b question-only baselines.

Provenance & license

• Kvasir-SEG (Jha et al., MICCAI 2020, CC BY 4.0) — 345 images in the benchmark subset.
• CVC-ClinicDB (Bernal et al., 2015) — 283 images.
• ETIS-LaribPolypDB (Silva et al., 2014) — 132 organic + 133 composite images.
• Questions, answers, programs, traces, and verified labels were produced by the SegVQA pipeline (DSL sampling → LLM paraphrase → independent symbolic verifier). Composite images are synthetic multi-lesion composites for quota filling; the artifact detector reports 0.944 accuracy distinguishing composited vs. organic frames (transparency check).

Quick start (other-server eval)

from datasets import load_dataset
ds = load_dataset("AiventraLab/segvqa-benchmark", split="train")
# image for item i:
img = ds[i]["file_name"]
# to load the actual image:
from PIL import Image
im = Image.open(f"https://huggingface.co/datasets/AiventraLab/segvqa-benchmark/resolve/main/{img}")

Or git lfs pull / hf download AiventraLab/segvqa-benchmark to mirror the whole dataset.