--- dataset_info: features: - name: images list: image: decode: false - name: image_group_id dtype: string - name: qa list: - name: inputs dtype: string - name: choices_list list: string - name: labels dtype: int64 - name: metadata dtype: string splits: - name: validation num_bytes: 240432366 num_examples: 2000 - name: train num_bytes: 12212389445 num_examples: 100000 download_size: 12453782687 dataset_size: 12452821811 configs: - config_name: default data_files: - split: validation path: data/validation-* - split: train path: data/train-* language: - en license: cc-by-4.0 source_datasets: - maveryn/trace task_categories: - multiple-choice - visual-question-answering tags: - decision-model - procedural --- # trace-mcq TRACE MC retains native training/validation QAs and adds fresh training generations. Original data: https://huggingface.co/datasets/maveryn/trace. New instances use the pinned TRACE generator and independent seeds; native training answers supply same-task distractors checked with the original scorer. The native validation split is preserved, and image/question/gold duplicates and held-out image copies are excluded from the fresh training additions. Seeds and rendering versions are recorded; execution traces and gold targets stay out of model inputs. Dataset terms: CC BY 4.0, with the original asset credits and licenses retained. Original data: [maveryn/trace](https://huggingface.co/datasets/maveryn/trace), [original distribution](https://huggingface.co/datasets/maveryn/trace), [upstream generator and verifier](https://github.com/maveryn/trace). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/procedural/trace_generate.py](https://github.com/sileod/tasksource/blob/main/scripts/repackage_dataset/procedural/trace_generate.py#L75). [Conversion provenance](provenance.json): source revisions, hashes, configuration and exclusions. [Upstream asset credits and licenses](asset-notices.json) accompany the original rendered images.