Missing checkpoint and merge script for LLaVA-OneVision-2-stage-0 (initial unaligned model)

[foamliu]

Description:

Issue Summary

In LLaVA-OneVision-1.5, the README provided a "Checkpoint and Format Conversion" section with two options to obtain the initial (unaligned) multi-modal model (stage-0):

1. Download pre-trained LLaVA-OneVision-1.5-4B-stage0 from Hugging Face
2. Merge the original ViT and LLM weights using a provided script (ds/merge_model.py)

However, for LLaVA-OneVision-2, I cannot find:

• A similar merge script for combining the vision encoder and LLM
• The stage-0 checkpoint on Hugging Face (e.g., LLaVA-OneVision-2-4B-stage0)

Questions

1. Does LLaVA-OneVision-2 have a similar merge script to obtain the initial unaligned model? If so, where can I find it?

2. If not, is there a pre-trained stage-0 checkpoint available for download? I searched the Hugging Face repo but couldn't find it.

What I'm trying to do

I want to start from the initial merged (but not yet aligned) multi-modal model to train on my own custom data, similar to how stage-0 worked in v1.5.

LLaVA-OneVision-1.5 reference (working example)

From the v1.5 README:

python ds/merge_model.py \
--vit_path DeepGlint-AI/rice-vit-large-patch14-560 \
--llm_path Qwen/Qwen3-4B-Instruct-2507 \
--output LLaVA-OneVision-1.5-4B-stage0

Environment

• Model: LLaVA-OneVision-2
• Looking for equivalent: stage-0 checkpoint or merge script

---

Additional context: I understand that v2 may have different architecture (possibly different ViT/LLM combination), but I'm hoping there's an equivalent starting point for training from the merged but unaligned stage.

Thank you for your great work on this model!