cis-lmu/uniOCR.bench-v1.0-results

Document OCR using gemini-3.1-flash-lite-preview

OCR results from cis-lmu/uniOCR.bench using gemini-3.1-flash-lite-preview via the native Gemini API.

Processing Details

Field	Value
Source Dataset	cis-lmu/uniOCR.bench
Model	gemini-3.1-flash-lite-preview
Samples	16,375
Processing Time	452.8 min
Date	2026-03-15 10:01 UTC
Image Column	img_plain
Output Column	markdown
Split	test
Max Output Tokens	8,192
Workers	1
Thinking	minimal (thinking_level=minimal)

Supported Models

Model	API String	Thinking
Gemini 2.0 Flash	gemini-2.0-flash	none
Gemini 2.5 Flash-Lite	gemini-2.5-flash-lite	disabled
Gemini 2.5 Flash	gemini-2.5-flash	disabled
Gemini 3 Flash	gemini-3-flash-preview	minimal
Gemini 3.1 Flash-Lite	gemini-3.1-flash-lite-preview	minimal
Gemini 3.1 Pro	gemini-3.1-pro-preview	minimal

Reproduction

uv run gemini-ocr.py \
    cis-lmu/uniOCR.bench \
    <output-dataset> \
    --model gemini-3.1-flash-lite-preview \
    --image-column img_plain \
    --max-tokens 8192 \
    --workers 1