eval-code-slidevqa-mathvista-mathverse-20260614

# SlideVQA / MathVista / MathVerse Evaluation Code This bundle collects the evaluation code currently used on this server for three datasets: - `slidevqa/` - `mathvista/` - `mathverse/` ## Files - `slidevqa/eval_slidevqa_f1.py` - Source: `/local_data/pengchunyi/eval_slidevqa_f1.py` - Purpose: evaluates SlideVQA predictions with exact match and token-level F1. - `mathvista/eval_mathvista_vlmevalkit_xlsx.py` - Source: `/local_data/pengchunyi/Mopd/mopd_private_bundle_no_domain5_20260610/code/VL_Benchmarks_4Suite_CodeOnly/eval_mathvista_vlmevalkit_xlsx.py` - Purpose: evaluates MathVista prediction exports with the local VLMEvalKit-style heuristic. - `mathverse/mathverse.py` - Source: `/local_data/pengchunyi/VLMEvalKit/vlmeval/dataset/utils/mathverse.py` - Purpose: MathVerse evaluation utilities used by VLMEvalKit, including answer extraction, scoring, and aggregated accuracy. - Note: this file depends on `vlmeval.smp.file.load` from VLMEvalKit and may also require a judge model / API depending on how it is used. - `mathverse/eval_mathverse_acc.py` - Source logic: extracted from `/local_data/pengchunyi/Mopd/VL_Benchmarks_4Suite_CodeOnly/unified_vllm_inference.py` - Purpose: standalone local MathVerse accuracy recomputation script using the patched `parse_choice_answer` and `freeform_answers_match` logic that was used for local recomputed summaries. - `upload_eval_code_to_modelscope.py` - Purpose: uploads this bundle file-by-file to a ModelScope dataset repository. ## Notes - This bundle preserves the exact source files rather than rewriting them. - `MathVerse` does not currently live here as a single standalone script on this machine; the active evaluation logic is the VLMEvalKit utility file included above. - If needed later, this bundle can be extended with a standalone wrapper for `MathVerse`.