Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,48 @@
+---
+license: apache-2.0
+task_categories:
+  - other
+tags:
+  - motion-forecasting
+  - autonomous-driving
+  - waymo
+  - motionlm
+  - trajectory-prediction
+pretty_name: MotionLM Processed WOMD Shards
+size_categories:
+  - 1K<n<10K
+---
+
+# MotionLM Processed WOMD Dataset
+
+A processed, lightweight version of the [Waymo Open Motion Dataset (WOMD)](https://waymo.com/open/data/motion/), prepared for training a [MotionLM](https://github.com/Mol2017/MotionLM) model.
+
+The original WOMD scenarios have been transformed into compact PyTorch tensor shards (`.pt.zst`, zstd-compressed) so they can be streamed directly into the training loop without the overhead of decoding protobuf scenarios on the fly.
+
+## Splits
+
+| Split        | Shards | Size  |
+| ------------ | ------ | ----- |
+| `training`   | 1000   | ~114 GB |
+| `validation` | 150    | ~10 GB  |
+| `test`       | 150    | ~11 GB  |
+
+File naming follows `{split}.{i}of{n}.pt.zst`.
+
+## Usage
+
+Used to train the MotionLM implementation at https://github.com/Mol2017/MotionLM. See that repository for the data loader, model code, and training scripts.
+
+```python
+import torch
+import zstandard as zstd
+import io
+
+with open("training/train.0of1000.pt.zst", "rb") as f:
+    decompressed = zstd.ZstdDecompressor().decompress(f.read())
+batch = torch.load(io.BytesIO(decompressed))
+```
+
+## Source
+
+Derived from the Waymo Open Motion Dataset. Please consult the [Waymo Open Dataset license](https://waymo.com/open/terms/) for terms governing use of the underlying data.