saketh-chervu/reuse-vs-recraft-slice150

The 150 conflict slices — the reuse-vs-recraft repair/damage test set

The standing conflict-slice metric of the reuse-vs-recraft programme (program-wide fix, 2026-08-16). 150 items: for each of three benchmarks, 25 repair-side items the untrained model (Qwen2.5-VL-7B-Instruct, greedy, 48 new tokens) answers WRONGLY — it follows the misleading text — and 25 cost-side items it answers CORRECTLY — it resists. A trained model is scored on both sides:

• fixed — repair-side items it now answers correctly (of 25)
• broken — cost-side items it now answers wrongly (of 25)
• reported per benchmark, exactly 3 numbers: net percentage (= fixed% − broken%), then (fixed / broken) raw counts. Never pooled raw across benchmarks.

The earlier 75-item version measured only the repair side and was structurally blind to cost; that is why this set exists.

The three benchmarks

benchmark	what the misleading text is	source
blind_faith	a false single-turn text description beside a document photograph	dal-289/word_or_vision (DocVQA pictures)
mmhal_snowball	a hallucinated claim built up over prior conversation turns (the history is part of the item)	MM-Hallu/MMHalSnowball
phd	a paragraph asserting a wrong fact about a natural photograph, followed by the benchmark's fixed "follow the image" instruction	AIMClab-RUC/PhD

Files

• items.jsonl — one line per item: item_id, benchmark, side (repair / cost), gold, options, meta, and request (kind, system, history as a list of {role, text, image_files} turns, final text, image_files). The prompt fields are verbatim what the model receives; nothing is re-rendered.
• images/<benchmark>/<side>/… — every image, original encoding kept.
• ids.json — the 150 item ids grouped benchmark → side, verbatim from the frozen selection.

Provenance

Selection frozen 2026-08-06, seed 0, deterministic and stratified, from the untrained model's own baseline records (slices.json groups repair and damage_conflict in saketh-chervu/reuse-vs-recraft-substrate, which also carries these items in the programme's runnable pickle form under slice_bundle/). Verified 2026-08-16: the untrained model is wrong on 25/25/25 of the repair side and right on 25/25/25 of the cost side in the banked baseline records. Items derive from the three public benchmarks named above; their original licenses apply to the underlying images and questions.