Open Models Almanac
Methodspost-trainingsupervised fine-tuning

single mention · post-training

Sandwich fine-tuning

Sandwich fine-tuning is a supervised fine-tuning strategy for VLA policies that fine-tunes the vision encoder, token embeddings, and final layer while leaving the LLM backbone frozen.1 It is motivated by findings that fine-tuning only the last layer or freezing the vision encoder performs poorly, so the visual features need adaptation to the target scene.2

Trade-off It achieves better performance than partial fine-tuning while using less GPU memory than fine-tuning the full backbone.3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. it fine-tunes the vision encoder, and it consumes less GPU memory since it does not fine-tune the full LLM backbone — OpenVLA: An Open-Source Vision-Language-Action Model
  2. only fine-tuning the network’s last layer or freezing the vision encoder leads to poor performance — OpenVLA: An Open-Source Vision-Language-Action Model
  3. sandwich fine-tuning” achieves better performance since it fine-tunes the vision encoder, and it consumes less GPU memory — OpenVLA: An Open-Source Vision-Language-Action Model

Filed under post-training › supervised fine-tuning