Tech Graphics Benchmark Runs
Tech Graphics Benchmark: generations Write-up: Can image models spell the labels? A first look at my Tech Graphics Benchmark Every image generated for Tech-Graphics-Benchmark, one folder per round. Prompts: danielrosehill/Tech-Graphics-Benchmark Annotations (raw eval data): danielrosehill/Tech-Graphics-Benchmark-Ratings Layout: <run_id>/manifest.jsonl + <run_id>/images/<gen_id>.png. Manifest rows are shuffled and gen_ids are opaque, so the annotation Space stays blind.… See the full description on the dataset page: https://huggingface.co/datasets/danielrosehill/Tech-Graphics-Benchmark-Runs.
About This Project
Tech-Graphics-Benchmark-Runs serves as a centralized repository for image generation outputs derived from the Tech-Graphics-Benchmark prompts. It organizes data by individual runs, providing a systematic structure where each directory contains a manifest file and corresponding image assets to support rigorous evaluation of text-to-image and image-to-image models.
The project employs a blind evaluation methodology by utilizing shuffled manifest rows and opaque generation identifiers, ensuring that annotation processes remain unbiased. It is designed for researchers and developers working on model benchmarking, specifically those focusing on text rendering and technical diagram synthesis. The dataset is distributed under a CC-BY-4.0 license and integrates with ML Croissant for enhanced interoperability with machine learning workflows.