Add dataset card with documentation

README.md

@@ -0,0 +1,143 @@
+---
+license: mit
+task_categories:
+  - question-answering
+  - text-generation
+language:
+  - en
+tags:
+  - finance
+  - financial-qa
+  - benchmark
+  - evaluation
+size_categories:
+  - 1B<n<10B
+---
+
+# FinBench Datasets
+
+Financial question-answering benchmark datasets for evaluating LLMs on financial reasoning tasks.
+
+## Datasets Included
+
+| Dataset | Description | Size | Files |
+|---------|-------------|------|-------|
+| **docfinqa** | SEC filings analysis - long document QA | ~5GB | train.json, dev.json, test.json |
+| **finqa** | Financial reports with tables | ~9MB | test.json |
+| **tatqa** | Table and text QA | ~17MB | train, dev, test splits |
+| **docmath_eval** | Document math problems | ~283MB | 8 complexity splits |
+
+## Quick Start
+
+### Download All Datasets
+
+```bash
+# Install huggingface_hub
+pip install huggingface_hub
+
+# Download to data/ folder
+huggingface-cli download Ayushnangia/finbench-data --local-dir data --repo-type dataset
+```
+
+### Python Download
+
+```python
+from huggingface_hub import snapshot_download
+
+snapshot_download(
+    "Ayushnangia/finbench-data",
+    repo_type="dataset",
+    local_dir="data",
+    local_dir_use_symlinks=False  # For Windows compatibility
+)
+```
+
+## Dataset Structure
+
+```
+data/
+├── docfinqa/
+│   ├── train.json
+│   ├── dev.json
+│   ├── test.json
+│   └── docfinqa.json/
+│       ├── train.json
+│       ├── validation.json
+│       └── test.json
+│
+├── finqa/
+│   └── test.json
+│
+├── tatqa/
+│   ├── tatqa_dataset_train.json
+│   ├── tatqa_dataset_dev.json
+│   ├── tatqa_dataset_test.json
+│   └── tatqa_dataset_test_gold.json
+│
+└── docmath_eval/
+    └── docmath_eval.json/
+        ├── complong_test.json
+        ├── complong_testmini.json
+        ├── compshort_test.json
+        ├── compshort_testmini.json
+        ├── simplong_test.json
+        ├── simplong_testmini.json
+        ├── simpshort_test.json
+        └── simpshort_testmini.json
+```
+
+## Dataset Details
+
+### DocFinQA
+- **Source**: SEC financial filings
+- **Task**: Long-document question answering
+- **Context**: Full financial documents (~100k tokens)
+- **Use case**: Testing LLM ability to reason over lengthy financial documents
+
+### FinQA
+- **Source**: Financial reports with tables
+- **Task**: Numerical reasoning with tables
+- **Use case**: Testing arithmetic and table understanding
+
+### TAT-QA
+- **Source**: Financial reports (tables + text)
+- **Task**: Hybrid table-text QA
+- **Use case**: Testing combined reasoning over structured and unstructured data
+
+### DocMath-Eval
+- **Source**: Mathematical documents
+- **Task**: Document-level math problems
+- **Splits**: By complexity (simple/complex) and length (short/long)
+- **Use case**: Testing mathematical reasoning in document context
+
+## Usage with FinBench
+
+These datasets are designed for use with the [FinBench](https://github.com/ayushnangia/finbench) evaluation framework:
+
+```bash
+# Clone FinBench
+git clone -b ananya-setup https://github.com/ayushnangia/finbench.git
+cd finbench
+
+# Setup (Windows)
+.\setup-windows.ps1
+
+# Download datasets
+huggingface-cli download Ayushnangia/finbench-data --local-dir data --repo-type dataset
+
+# Run evaluation
+python finbench.py run docfinqa_vibe_test --limit 10
+```
+
+## Citation
+
+If you use these datasets, please cite the original sources:
+
+- **DocFinQA**: [Kensho DocFinQA](https://huggingface.co/datasets/kensho/DocFinQA)
+- **FinQA**: [ChanceFocus FinQA](https://huggingface.co/datasets/ChanceFocus/finqa)
+- **TAT-QA**: [TAT-QA Dataset](https://github.com/NExTplusplus/TAT-QA)
+- **DocMath-Eval**: [Yale NLP DocMath-Eval](https://huggingface.co/datasets/yale-nlp/DocMath-Eval)
+
+## License
+
+MIT License - See individual dataset licenses for specific terms.