Add hipho physics dataset (#1318)

* Initial commit fot HiPhO

* update

* update

* refactor: simplify HiPhO dataset logging system

- Remove custom LogBuffer class and thread-safe logging
- Replace safe_print with standard print statements
- Remove threading and datetime imports
- Simplify build_prompt function by removing verbose debug output
- Update dataset URL from haiyuanwan/HiPhO to HY-Wan/HiPhO
- Reduce code from 899 to 803 lines (10.7% reduction)
- Maintain all core functionality: evaluation logic, prompt building, hipho_verifier integration

* refactor: remove parallel evaluation framework from HiPhO dataset

- Remove complex parallel evaluation using track_progress_rich
- Simplify to sequential evaluation for better stability and debugging
- Remove multiprocessing and parallel task management dependencies
- Rename functions to remove '_with_buffer' suffix and log_buffer parameters
- Remove nproc parameter handling and temporary file management
- Reduce code from 803 to 774 lines (additional 3.6% reduction)
- Maintain all core evaluation logic: fine/coarse-grained scoring, hipho_verifier integration
- Sequential evaluation is sufficient for physics olympiad problem counts

* refactor: major simplification of HiPhO dataset implementation

Major improvements:
- Remove 6 unnecessary try-except blocks that were hiding errors
- Standardize judge model initialization to follow VLMEvalKit conventions
- Move all prompt templates to utils/prompt_inference.py for better organization
- Remove redundant count statistics (fine_grained_count, coarse_grained_count, total_count)
- Remove unused fallback functions (_simple_answer_matching, _extract_prediction_for_display)
- Fix multi-image base64 processing bug
- Correct dataset name display in summary output
- Remove verbose debugging output and unnecessary comments

Code reduction: 899 → 604 lines (32.8% reduction)
Eliminated potential bugs and improved maintainability while preserving all core functionality

* Improve HiPhO dataset: translate comments to English and enhance configuration

- Translate all Chinese comments to English in hipho.py, hipho_verifier.py, and prompt_inference.py
- Simplify comments while maintaining technical accuracy
- Replace hardcoded verifier model configuration with environment variables
- Use VLMEvalKit standard environment variable approach for better flexibility
- Add support for HIPHO_VERIFIER_* environment variables for model configuration
- Improve code maintainability and international accessibility

* Add new dependencies for HiPhO dataset functionality

- Add datasets: for HuggingFace dataset loading
- Add scikit-learn: for machine learning utilities
- Add pylatexenc==2.10: for LaTeX text processing
- Add math-verify: for mathematical answer verification

These dependencies are required for the HiPhO physics olympiad dataset
evaluation and verification functionality.

* Add hipho_prompt_inference.py utility file

* Update import statement for prompt inference module

---------

Co-authored-by: Haodong Duan <[email protected]>
Co-authored-by: Ma Zerun <[email protected]>

Author: 3 people

requirements.txt

@@ -1,4 +1,5 @@
 accelerate
+datasets
 dotenv
 einops
 # for gemini api
@@ -8,6 +9,7 @@ huggingface_hub
 imageio
 ipdb
 json_repair
+math-verify
 matplotlib
 nltk
 numpy
@@ -19,10 +21,12 @@ pandas
 pillow
 portalocker
 protobuf
+pylatexenc==2.10
 python-dotenv
 qwen_vl_utils
 requests
 rich
+scikit-learn
 sentencepiece
 setuptools
 sty

vlmeval/dataset/__init__.py

@@ -93,6 +93,7 @@
 from .matbench import MATBench
 
 from .reasonmap_plus import ReasonMap_Plus
+from .hipho import HiPhODataset
 from .gsm8k_v import GSM8KVDataset
 
 
@@ -223,7 +224,7 @@ def evaluate(self, eval_file, **judge_kwargs):
     OmniEarthMCQBench, VisFactor, OSTDataset, OCRBench_v2, TreeBench, CVQA, M4Bench,
     AyaVisionBench, TopViewRS, VLMBias, MMHELIX, MedqbenchMCQDataset, MathCanvas,
     MedqbenchPairedDescriptionDataset, MedqbenchCaptionDataset, ChartMuseum, ChartQAPro, ReasonMap_Plus,
-    olmOCRBench, OceanOCRBench, MATBench, VLRMBench, RefCOCODataset, SimpleVQA
+    olmOCRBench, OceanOCRBench, MATBench, VLRMBench, RefCOCODataset, SimpleVQA, HiPhODataset
 ]
 
 VIDEO_DATASET = [