Add base / coverage_aware / combined world-model checkpoints + model card

README.md

@@ -0,0 +1,68 @@
+---
+license: mit
+tags:
+- world-models
+- reinforcement-learning
+- robotics
+- video-prediction
+- dreamer
+---
+
+# MMBench2 World Model Checkpoints
+
+Pretrained and finetuned checkpoints for the 350M-parameter generative world model from the paper [**Hallucination in World Models is Predictable and Preventable**](https://www.nicklashansen.com/mmbench2) by [Nicklas Hansen](https://www.nicklashansen.com) and [Xiaolong Wang](https://xiaolonw.github.io) (UC San Diego).
+
+The world model follows the architecture and two-stage training recipe of [Dreamer 4](https://arxiv.org/abs/2509.24527), adapted for large-scale multi-task continuous control, and is trained on **MMBench2** — a 427-hour, 210-task dataset for visual world modeling (see the [dataset repository](https://huggingface.co/datasets/nicklashansen/mmbench2)). Each variant is a `(tokenizer.pt, dynamics.pt)` pair at 224×224 resolution:
+
+- **tokenizer** — a causal video tokenizer (≈50M-parameter encoder + ≈50M-parameter decoder, projecting to a 64-dim tanh-bounded latent).
+- **dynamics** — a ≈250M-parameter block-causal Transformer trained on the frozen tokenizer with a shortcut flow-matching objective, with reward-prediction and behavior-cloning heads.
+
+## Variants
+
+| Variant | Description |
+|---------|-------------|
+| `base` | Pretrained world model |
+| `coverage_aware` | Coverage-aware finetuned world model |
+| `combined` | Finetuned with all targeted data-collection sources |
+
+## Repository layout
+
+```
+base/            tokenizer.pt  dynamics.pt
+coverage_aware/  tokenizer.pt  dynamics.pt
+combined/        tokenizer.pt  dynamics.pt
+```
+
+All checkpoints are weights-only.
+
+## Usage
+
+Using the accompanying code release:
+
+```
+cd dreamer4
+python download_checkpoints.py --variant combined     # or: base | coverage_aware | all
+./run_interactive.sh combined                          # launch the interactive interface
+```
+
+`download_checkpoints.py` fetches the `(tokenizer.pt, dynamics.pt)` pair into `./checkpoints/<variant>/`. Alternatively, download directly with the Hugging Face CLI:
+
+```
+hf download nicklashansen/mmbench2-checkpoints --include "combined/*" --local-dir ./checkpoints
+```
+
+See the [paper](https://www.nicklashansen.com/mmbench2) and the code release for architecture details, training recipes, and the hallucination detection and mitigation methods.
+
+## License
+
+Released under the MIT License.
+
+## Citation
+
+```bibtex
+@article{Hansen2026Hallucination,
+    title={Hallucination in World Models is Predictable and Preventable},
+    author={Nicklas Hansen and Xiaolong Wang},
+    year={2026},
+}
+```

base/dynamics.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec28ba7dd40113d1b0e0debfc37df42de0616b39f62dcb01d7ce7431c8803b59
+size 1017872707

base/tokenizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95e521d36cba21bb47b14e4df4f1d626161c3f8f9955b75443f6b9b0b08fe323
+size 406957899

combined/dynamics.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a3c359af4447a84a29be398254437160dcbf3b747ad0c834a57281a5bf44d2
+size 1017872835

combined/tokenizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6d46bfd019d176167b9fd1a04aa77f6052c81d3a4dbe455776208f656f53ae1
+size 406958091

coverage_aware/dynamics.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb5316bc99476d30666c2d92068de2f040bcc3e9e8f873f3602a208ae6a1c3eb
+size 1017872707

coverage_aware/tokenizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a88c0ed422aa61947e17f57b639538f1c6218e67e682803db13da43996183bf2
+size 406958027