willow-alpha/vision_adapter

license

license
mit

Forge-1V Micro Vision Adapter Scaffold

This folder contains an experimental, untrained micro vision-adapter scaffold for future Forge-1V work.

It is intentionally separate from the main text checkpoint:

• The main config.json remains Llama-compatible.
• The GGUF export remains text-only.
• These files do not make the released model able to view images.

Suggested target design:

• Tiny patch encoder: 3-channel images to a small vision width.
• Projection: vision width to the 1024-dimensional Forge text hidden size.
• Prefix tokens: projected visual tokens can be prepended to the text sequence in a future custom multimodal training run.

Approximate extra parameters for the scaffold design are well under 1M, keeping the total system under 400M parameters.