20 days ago
-
Add October 2026 model results to leaderboard
Adds the new rows from the canonical GitHub leaderboard (mts-ai/MWS-Vision-Bench@461b9de). Existing rows are unchanged.
-
Update Cotype Pro 3 validation results
-
Add Gemini 3.6 Flash and GPT-5.5 validation results
-
Complete anti-fraud validation leaderboard
-
Add anti-fraud as a separate leaderboard metric
-
Remove leaderboard footnotes
-
Update validation leaderboard
-
Add Kimi K2 Instruct and Gemma 4 31B IT results
-
Add GLM-5V Turbo validation results
-
Add Qwen3.6 35B-A3B validation results
-
Normalize Cotype model naming
-
Limit leaderboard emphasis to top five models
-
Sync leaderboard with GitHub
-
Sync metadata with current parquet annotations
-
Добавить результаты Qwen3.6-27B и Gemini-3.1 Flash Lite в лидерборд
-
Add language config usage examples to README
-
Add zh/train from part1_ru_current_zh.json
-
Add en/train from part1_ru_current_en.json
-
Add en/train from part1_en.json
-
Update README.md
-
Update README.md
-
Update README.md
Small style fix -
Update README.md
Updated leaderboard: fixed VQA category + new models in the leaderboard.
-
Update reasoning VQA Q/A (train)
-
Leaderboard update
-
Add leaderboard with Qwen3-VL models
-
Add Qwen3-VL-4B and Qwen3-VL-8B results to leaderboard
-
Add Pixtral-12B-2409 results to leaderboard
-
Update README.md
-
Update README.md
-
Delete prev.png
-
Add preview image section to README
-
Add preview image with data examples
-
Add preview image with data examples
-
Update README.md
-
Update README: correct metadata.jsonl structure with file_name field
-
Fix README: annotations.json → metadata.jsonl
-
Update README with new statistics (removed screenshots)
-
Reupload dataset with metadata.jsonl format
-
initial commit