Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,77 @@
+---
+dataset_info:
+  features:
+    - name: id
+      dtype: int64
+    - name: name
+      dtype: string
+    - name: video_path
+      dtype: string
+    - name: count
+      sequence:
+        dtype: int64
+    - name: fuzzy_action
+      dtype: bool
+    - name: complex_action
+      dtype: bool
+  splits:
+    - name: test
+      num_examples: 227
+configs:
+  - config_name: default
+    data_files:
+      - split: test
+        path: data/test.jsonl
+license: cc-by-4.0
+task_categories:
+  - video-classification
+  - visual-question-answering
+tags:
+  - video
+  - counting
+  - repetition-counting
+  - exercise
+  - benchmark
+pretty_name: PushUpBench
+size_categories:
+  - n<1K
+---
+
+# PushUpBench: Video Repetition Counting Benchmark
+
+PushUpBench is a benchmark for evaluating vision-language models on their ability to count exercise repetitions in videos.
+
+## Dataset Description
+
+- **Total samples**: 227
+- **Video format**: MP4
+- **Task**: Count the number of repetitions of a specified exercise in a video
+
+## Dataset Structure
+
+Each sample contains:
+- `name`: Action description (e.g., "push ups", "leg lift", "knee to chest")
+- `video_path`: Filename of the video
+- `count`: List of acceptable count values (some exercises have ambiguous boundaries)
+- `fuzzy_action`: Whether the action has ambiguous start/end boundaries
+- `complex_action`: Whether the action is compound/complex
+
+## Usage with lmms-eval
+
+```bash
+# Set the video directory
+export PUSHUPBENCH_VIDEO_DIR=/path/to/videos
+
+# Run evaluation
+python -m lmms_eval \
+    --model <model> \
+    --tasks pushupbench \
+    --batch_size 1 \
+    --output_path results/
+```
+
+## Metrics
+
+- **Exact Match**: Prediction matches any value in the ground truth count list
+- **MAE**: Mean Absolute Error between prediction and primary ground truth
+- **OBO**: Off-By-One accuracy (prediction within 1 of any ground truth)