---
license: apache-2.0
base_model: BestWishYsh/Helios-Distilled
library_name: peft
pipeline_tag: text-to-video
inference: false
tags:
  - worl
  - helios
  - video-generation
  - world-model
  - reinforcement-learning
  - lora
  - safetensors
---

# WORL-Helios

**WORL-Helios** is a reinforcement-learning-trained LoRA adapter for
[Helios-Distilled](https://huggingface.co/BestWishYsh/Helios-Distilled).
It adapts the video transformer for autoregressive video generation.

<img src="assets/worl-framework.png" alt="WORL video reinforcement-learning framework" width="100%">

Related releases: [WORL-NVRM](https://huggingface.co/Efficient-Large-Model/WORL-NVRM)
and [WORL-Lingbot](https://huggingface.co/Efficient-Large-Model/WORL-Lingbot).

## Model details

| Field | Description |
| --- | --- |
| Base model | `BestWishYsh/Helios-Distilled` |
| Transformer class | `HeliosTransformer3DModel` |
| Released artifact | Current-policy PEFT LoRA adapter |
| Adapter precision | FP32 |
| LoRA rank / alpha | 128 / 128 |
| Attention targets | `to_q`, `to_k`, `to_v`, `to_out.0` |
| Coverage | 40 transformer blocks; 640 adapter tensors |
| Adapter parameters | 419,430,400 |
| Input / output | Text and video/history conditioning through the Helios runtime / generated video continuation |

This repository contains an adapter, not a complete video-generation pipeline.
The base transformer, text encoder, tokenizer, VAE, scheduler and compatible
Helios runtime are required separately.

## Training data and method

The released adapter's export record specifies WORL training with VideoAlign
VQ, MQ and TA rewards, interactive prompt mode and 16 FPS. This description
follows the actual exported artifact. The exact prompt corpus and its size are
not documented in that record, and no prompt dataset is bundled here.

## Download and load

```bash
hf download Efficient-Large-Model/WORL-Helios --local-dir artifacts/WORL-Helios
```

With [Helios](https://github.com/PKU-YuanGroup/Helios) and a compatible PEFT
environment installed, attach the adapter to the base transformer:

```python
import torch
from peft import PeftModel
from helios.modules.transformer_helios import HeliosTransformer3DModel

base = HeliosTransformer3DModel.from_pretrained(
    "BestWishYsh/Helios-Distilled",
    subfolder="transformer",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
).eval()

model = PeftModel.from_pretrained(
    base,
    "Efficient-Large-Model/WORL-Helios",
    is_trainable=False,
).eval()
```

This loads the transformer and adapter. Video generation additionally needs
the remaining Helios pipeline components. The source configuration used
`Wan-AI/Wan2.1-T2V-14B-Diffusers` for its VAE and records PEFT 0.18.1.
The exact Helios base revision was not retained in the source export; pin a
compatible base revision when reproducing an inference setup.

## Files and provenance

| Path | Contents |
| --- | --- |
| `adapter_model.safetensors` | FP32 policy adapter, approximately 1.68 GB |
| `adapter_config.json` | PEFT configuration |
| `RELEASE_MANIFEST.json` | Source revision and SHA-256 file inventory |
| `LICENSE.txt`, `NOTICE.txt` | License and source attribution |

Weights and adapter configuration are unchanged from
[qyoo/helios-worl at c17176f56773](https://huggingface.co/qyoo/helios-worl/tree/c17176f5677301f9e10abb8d0ff69733e60b5e6d).
Documentation and release metadata were updated for this repository. The
original export record remains available at that source revision.

File integrity is checked against the source snapshot. This publication does
not establish a new generation benchmark or hardware compatibility matrix.

## Intended use and limitations

The adapter supports research on video generation and reinforcement learning
of world models. Long generations can accumulate temporal artifacts, identity
drift, geometric inconsistency and prompt mismatch. Generated video is
synthetic and is not a validated physical simulation. Review generated outputs
for the intended use and respect rights in input media.

## License

This release retains the source model card's
[Apache License 2.0](LICENSE.txt). The separately obtained Helios base model
and runtime components retain their upstream licenses and notices. See
[NOTICE.txt](NOTICE.txt) for source attribution.