helios-worl

<div align="center">

Helios WoRL

A Helios-Distilled policy adapter trained with the All-in-One RL Framework of Video World Model

WoRL · Base model · Helios paper · Helios code

</div>

Helios WoRL is a current-policy PEFT LoRA export from a WoRL run built on BestWishYsh/Helios-Distilled. It contains the inference adapter and its configuration; it is not a standalone video pipeline.

Attach the adapter to the matching Helios-Distilled transformer in a compatible Helios runtime. The text encoder, tokenizer, scheduler, VAE, and other pipeline components are not duplicated here.

WoRL in one figure

<p align="center"> <img src="assets/worl-framework.png" width="100%" alt="WoRL reinforcement-learning framework"> </p>

WoRL combines autoregressive rollout context, group generation, video reward, and policy optimization for long-horizon video models. This archived Helios run was configured with VideoAlign reward dimensions VQ, MQ, and TA.

Release overview

Repository layout

The optimizer, EMA state, behavior-policy snapshot, reference adapter, and training-state pickle are intentionally excluded from this inference release. They remain in the internal training-resume archive.

Load the adapter

Use the Helios implementation that provides helios.modules.transformer_helios.HeliosTransformer3DModel, then attach the adapter with PEFT:

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,
    "qyoo/helios-worl",
    is_trainable=False,
).eval()

For full generation, supply the remaining Helios pipeline components. The archived run used Wan-AI/Wan2.1-T2V-14B-Diffusers as its VAE source.

Configured training profile

These fields were read from the archived training state and its launcher/config bundle.

Export integrity

• Adapter SHA-256: 51028a3a9e8b3db60088fda7ad96ad143dae9756310fca19981f3dcbadc71957.
• Configuration SHA-256: d2d5a95aa83e84d1580f732fafaee17ebbe1286404a15cf7859c3bcec42e1257.
• The adapter hash was verified both after download and against an independent direct stream of the source object.
• The safetensors header reports 640 FP32 tensors and PyTorch-format metadata.

Provenance

The exact training configuration is preserved with the internal checkpoint. The closest committed LongNFT source capture is commit 6405cf82d1326e822187eec09983c5f4ca0ab275. The checkpoint was written on 2026-04-09 before that working tree was committed, and the launcher did not record the working tree's Git identity or dirty patch. The configuration and released adapter bytes are preserved exactly, but byte-for-byte reconstruction of the uncommitted training code cannot be established retrospectively.

Limitations and responsible use

Long autoregressive generations can accumulate temporal artifacts, identity drift, geometric inconsistency, prompt mismatch, and motion errors. Generated video is synthetic and should not be treated as factual simulation or used as an autonomous-control signal. Review outputs and apply safeguards appropriate to the deployment context.

License

This repository is released under Apache 2.0. Helios-Distilled and its runtime components remain subject to their upstream licenses and notices.