Spaces:
Sleeping
Sleeping
Commit History
-
Add Kaggle 80-class hard-mode benchmark for Qwen3-VL-4B q4
Ronti PatangeBare prompt: 39.6% (95/240). Disambiguated per-class prompt: 50.8% (122/240, +11.2 pts), same technique and gain magnitude as the 20-class benchmark. Also documents a Vulkan mmproj hang on oversized (>1024px) images on the Arc iGPU, fixed by downscaling before serving.e5ffef0 -
Replace approx-sign tildes with U+2248 to avoid GFM strikethrough misparse
Ronti Patangeb8085e5 -
Read KHANA_API_KEY from env var instead of hardcoding it
Ronti Patangefd7e0ff -
Add full VLM benchmark suite and results: model comparison, prompt engineering, ZeroGPU deploy
Ronti Patange84d7f71 -
Add RAW_PROMPT: mode to analyze() for direct benchmark comparison
Ronti Patange45aa0de -
Surface exception details in analyze() response for debugging
Ronti Patange347d0b0 -
Add torchvision (required by Qwen3-VL's video sub-processor)
Ronti Patange9e1643c -
Fix invalid gradio sdk_version, update README for Qwen3-VL-4B
Ronti Patangef9ea036 -
Upgrade KhanaVision Space to Qwen3-VL-4B-Instruct
Ronti Patange45898ed
-
Initial KhanaVision inference Space
Ronti Patangef741165