bmeivar/metric-vqa-qwen3-235b-bundle

Qwen3-VL-235B bundle (v2) — metric_vqa probes for RITS

Self-contained bundle for running Qwen3-VL-235B on every metric_vqa probe. Built 2026-05-17.

Setup

pip install requests pandas Pillow tqdm
export RITS_API_BASE="https://.../v1"
export RITS_API_KEY="..."

Run a single condition

python scripts/run_qwen235b.py \
    --manifest manifests/in_the_wild_vision.csv \
    --output-dir out/in_the_wild_vision

Output: predictions.csv + metrics.json in the output dir.

Run everything (sequential)

for m in manifests/*.csv; do
    name=$(basename "$m" .csv)
    python scripts/run_qwen235b.py --manifest "$m" --output-dir "out/$name"
done

Conditions

Manifest	Rows	What it tests
in_the_wild_vision.csv	331	Baseline: full image + question (v3 typo-fixed).
in_the_wild_gray.csv	331	Image replaced with a uniform gray (127/255). Tests if the model attends to vision tokens that carry no information.
in_the_wild_text_only.csv	331	NO image attached. Pure LM-prior probe — bypasses vision encoder entirely.
in_the_wild_zoom_in.csv	~342	Object-only image, target scaled 1.4×. Apparent-size probe.
in_the_wild_zoom_out.csv	~342	Object-only image, target scaled 0.6×.
in_the_wild_distorted_k015.csv	~336	Barrel fish-eye k=0.15, edge-preserved (no content loss). Vanishing-point probe.
in_the_wild_distorted_k025.csv	~336	Barrel k=0.25, stronger.
in_the_wild_distorted_pin015.csv	~336	Pincushion k=−0.15 (reverse).
in_the_wild_distorted_pin025.csv	~336	Pincushion k=−0.25.
in_the_wild_target_plus_ref.csv	1254	All (target, reference) composites. Tests explicit reference-object use.
in_the_wild_context_5variants.csv	~1655	5 variants (full/object_only/ring/wrong/random) × queries. Context-cue ablation.
objectron_vision_unique27.csv	27	Vision on the 9-category × 3-dimension grid. Broadcast to 10,482 in post.
objectron_gray_unique27.csv	27	Gray-blank version of the 27 grid queries.
objectron_text_only_unique27.csv	27	Text-only version of the 27 grid queries.

Notes

• All in-the-wild manifests use v3 queries (typos fixed, capitalization normalized). Objectron queries are templated and untouched.
• Manifests with image = "<text-only>" skip the image attachment entirely; the runner sends only the text prompt.
• For the Objectron 27-row probes, broadcast the 27 predictions back to the full 10,482-row Objectron CSV before computing per-instance metrics (predictions are deterministic given temp=0; one prediction per (category, query) is correct).
• Default instruction strings:
  • Vision probes: "Answer with a single float in centimeters."
  • Text-only probes: "Answer with the typical object size in centimeters as a single float."