TongtongLiang/natural-context-grounding-v2

Context Grounding Ordering — Two-stage 30

30 images, 60 grounding questions, exact contexts and annotations, Qwen3.8-27B reference outputs, and one detailed diagnostic case study. Updated 2026-10-23. Images are synthetic: 17 natural scenes and 13 charts/interfaces; eight charts share a three-panel layout with independent data.

Case study with image · Full protocol · Interactive 30-case report · Metrics

What this collection tests

Same image, background T, Q1 and Q2; change the order within the initial user message:

Order	Initial user message
A	Image → T → Q1
B	T → Image → Q1
C	T → Q1 → Image

Generate Q1, append that condition's actual final A1 as assistant, then append Q2 as user. Image appears once. Thinking is saved separately and excluded from A1. The image-only baseline I asks each question independently. Exact system and stimulus text are in benchmark.json; protocol.py constructs messages.

Cases were actively selected and confirmed for Q1: A/C correct, B wrong; Q2: A correct, B/C wrong. The selected set is useful for reproduction and diagnostic comparison; its success/failure proportions describe the selection, not a natural-task prevalence estimate.

Qwen reference results

Qwen/Qwen3.8-27B via Tinker, qwen3_8_medium_reasoning, medium thinking, temperature 0, seed 20261023, max output 8192. Confirmation results, macro point F1 with GT box edges expanded by 5 on the 0–1000 coordinate scale:

Target	Image-only I	A	B	C
Q1	100%	100%	35.23%	100%
Q2	100%	100%	37.48%	45.02%

Q1 exact cases: I/A/C 30/30, B 0/30. Q2 exact cases: I/A 30/30, B/C 0/30. Partial matches contribute to F1. Match predicted points to GT boxes one-to-one; F1 = 2 × matches / (predictions + GT). Strict scores use unexpanded boxes. Bounding boxes are preserved; selection uses point matching. metrics.json contains initial and confirmation results separately.

Simple reading: placing Q1 before the image can recover Q1 while leaving a later visual task vulnerable. Failures include choosing the wrong parent object, missing a group of parts, and coordinate drift. The n105 case study further shows that direct color recognition can succeed while part description or relation comparison assigns the target to a different object. Query wording can restore correct answers in that example.

Files and reproduction

• benchmark.json, images/, data/test.parquet: portable 30-case dataset. JSON is canonical; Parquet embeds image bytes for the viewer.
• reference_outputs.jsonl.gz: 480 reference record slots (30 × 8 conditions × initial/confirmation). Reused records carry provenance; this count is not a count of new paid calls.
• cases.json, source_evidence.jsonl.gz, source_manifests.json, prompts/: answers, source records and rendered input audits.
• diagnostics/: 60 completed A/C probe outputs on n105, including successful controls and confirmation calls; independent of the 30-case score.
• CASE_STUDY.md: n105 image, experiment and interpretation.
• protocol.py, scoring.py, validate.py: backend-neutral input construction and offline validation.

Download the repository and run python validate.py. For a new model, use protocol.py with the exact system from benchmark.json, evaluate all conditions, and feed its own A1 to Q2. Preserve content-block order and record renderer, reasoning mode and output budget. The report can be downloaded and opened locally.

from datasets import load_dataset
ds = load_dataset("TongtongLiang/natural-context-grounding-v2", split="test")
print(len(ds))  # 30

Version history

This release replaces the earlier Selected 20 at the repository head. That earlier version required Q1 correct in all three orders and included joint questions; its frozen revision remains 3e7ec57a90c054443a9f7751212b8babe0f5d821. The present 30-case set uses the sequential two-stage criterion above. No new inference was run for this publication.

CC BY 4.0. Attribute TongtongLiang, Context Grounding Ordering — Two-stage 30, and cite the repository revision used.