ysharma/north-micro-vision-pope-multilingual

Multilingual POPE hallucination eval: North-Micro-Vision-Instruct

POPE-style object-hallucination evaluation (POPE, Li et al. 2023, arXiv:2305.10355) of CohereLabs/North-Micro-Vision-Instruct (2.4B) in four languages, run on a T4 via HF Jobs.

Setup

• Questions: lmms-lab-encoder/POPE, config Full, split random — 100 COCO val2014 images, 3 questions per image (seeded sample of each image's 6), the same image+question in English, German, Spanish and Portuguese → 1,200 inferences (148 ground-truth "no" per language).
• Translation: hand-written template + noun translation table covering the 80 COCO categories POPE's random split draws from (e.g. Gibt es einen Hund im Bild?, ¿Hay un perro en la imagen?, Há um cachorro na imagem?), each ending with a same-shape "answer with only yes or no" instruction.
• Decoding: deterministic (do_sample=False), fp16 on t4-small, batch of 8, left padding, max 6 new tokens, transformers 5.16.0.
• Parsing: answers are parsed semantically across all four languages' yes/no words, because the model sometimes code-switches into English (e.g. answering "Yes." to a German question) — this is measured separately as the code-switching rate, not dropped.
• Metric: hallucination rate = fraction of ground-truth-"no" questions answered "yes" (POPE's false-positive / yes-hallucination rate).

Results

	English	German	Spanish	Portuguese
Hallucination rate (yes on "no")	1.35% (2/148)	1.35% (2/148)	1.35% (2/148)	5.41% (8/148)
Accuracy	89.3%	87.3%	81.3%	86.7%
False-negative rate (no on "yes")	19.7%	23.7%	35.5%	21.1%
Answer given in another language	0.0%	21.3%	8.0%	14.0%

Findings: object hallucination is near-floor in every language on this 100-image slice, but Portuguese hallucinates ~4× more than the others; Spanish shows the largest drop in accuracy (over-cautious "no" answers — 35.5% false negatives). German code-switches to English answers most often (21%).

Files

• summary.json — per-language metrics (confusion counts included)
• per_question.jsonl — all 1,200 records: image id, question in each language, ground truth, raw model answer, parsed answer
• eval.py — the exact evaluation script used
• Live metrics: ysharma/pope-multilingual-hallucination-trackio

Caveats

• Single run, deterministic decoding, one 100-image slice of the random split: rates on 148 negatives have wide confidence intervals (±~2-4% absolute); treat the de/es vs pt gap as directional.
• Prompt translations are hand-authored, not back-translated.