docs: add RIFT model card

README.md

@@ -0,0 +1,119 @@
+---
+license: apache-2.0
+library_name: fastwam
+pipeline_tag: robotics
+base_model: Wan-AI/Wan2.2-TI2V-5B
+base_model_relation: finetune
+datasets:
+  - yuanty/LIBERO-fastwam
+tags:
+  - robotics
+  - world-action-model
+  - flow-matching
+  - libero
+  - pytorch
+---
+
+# RIFT
+
+This private model repository contains the canonical RIFT checkpoint trained
+on LIBERO. RIFT is derived from
+[FastWAM](https://github.com/yuantianyuan01/FastWAM) at commit
+`45d8e1458921d83f8ad6cf9ce993d371208dabd0`.
+
+RIFT uses one representation-shaping flow-matching head. The checkpoint also
+retains a detached L2 diagnostic probe for schema compatibility; this probe
+does not shape the shared anticipation representation.
+
+## Files
+
+| File | Size | SHA-256 |
+|---|---:|---|
+| `rift_step021700.pt` | 12,060,942,777 bytes | `080e260b26c580cff5da0ce1de4c72e57eb2a6de47ae12f5712244dcf911f3cb` |
+| `dataset_stats.json` | 40,961 bytes | `f68a07ec322ea425164d94a8827aec1430109c7614637d03aec894aadc9fc05b` |
+
+`dataset_stats.json` is the training-run normalization file paired with this
+checkpoint. Keep both files in the same directory for automatic discovery.
+
+## Source and configuration
+
+- Source commit: [`184a582d7821e3a919dcd6e5795110ac0282eda6`](https://github.com/ChushanZhang/RIFT/tree/184a582d7821e3a919dcd6e5795110ac0282eda6)
+- Model configuration: [`config.yaml`](./config.yaml)
+- Canonical training task: [`configs/task/libero_rift_2cam224_1e-4.yaml`](https://github.com/ChushanZhang/RIFT/blob/184a582d7821e3a919dcd6e5795110ac0282eda6/configs/task/libero_rift_2cam224_1e-4.yaml)
+- Training step: 21,700
+- Checkpoint dtype: bfloat16
+
+The model depends on `Wan-AI/Wan2.2-TI2V-5B`, the tokenizer/text-encoder
+assets from `Wan-AI/Wan2.1-T2V-1.3B`, and the ActionDiT backbone generated by
+the preprocessing script in the source repository. The original training run
+did not record immutable Hub revisions for these external assets. Current
+dependency revisions must not be treated as proven training-time revisions.
+
+## Download
+
+```bash
+hf download PoopBear/RIFT \
+  rift_step021700.pt dataset_stats.json config.yaml \
+  --local-dir checkpoints/rift
+```
+
+Because this repository is private, authenticate with `hf auth login` first.
+
+Verify the artifacts before loading:
+
+```bash
+sha256sum checkpoints/rift/rift_step021700.pt \
+  checkpoints/rift/dataset_stats.json
+```
+
+## Install the matching source
+
+```bash
+git clone https://github.com/ChushanZhang/RIFT.git
+cd RIFT
+git checkout 184a582d7821e3a919dcd6e5795110ac0282eda6
+pip install -e .
+```
+
+The weight file contains `mot`, `proprio_encoder`, `step`, and `torch_dtype`.
+It is a weights-only checkpoint: optimizer, scheduler, and dataloader state are
+not included, so it is suitable for evaluation or warm-starting rather than an
+exact training resume.
+
+## Inputs and outputs
+
+The canonical LIBERO recipe uses two RGB cameras at 224 x 224, 8-dimensional
+proprioception, 7-dimensional actions, and a 32-step action horizon. Outputs
+must be interpreted using the included dataset statistics.
+
+## Evaluation status
+
+This artifact release does not claim a new benchmark result. Before reporting
+results, pin the simulator commit, task suite, seeds, episodes per task,
+action-chunking policy, and aggregation method.
+
+## Intended use and limitations
+
+The checkpoint is intended for research in controlled simulation. It is not
+intended for safety-critical control or unsupervised deployment on physical
+robots. It may fail under visual, linguistic, dynamics, or embodiment shift.
+Simulation performance does not establish real-world safety or robustness.
+
+PyTorch `.pt` files use pickle-based serialization. Load only the file whose
+checksum matches the value above, and prefer:
+
+```python
+torch.load(path, map_location="cpu", mmap=True, weights_only=True)
+```
+
+## License and attribution
+
+The checkpoint is distributed under Apache License 2.0, consistent with the
+Wan base model license. RIFT source code is separately distributed under the
+license in the source repository. See [`NOTICE`](./NOTICE) for provenance and
+attribution.
+
+## Citation
+
+Use the RIFT citation metadata from the matching source commit and also cite
+the upstream [Fast-WAM paper](https://arxiv.org/abs/2603.16666).