Releasing the Teacher Model (Human-Aligned SigLIP) for Custom AligNet Dataset Generation

Hello AligNet authors and contributors,

First, thank you for this excellent work and for open-sourcing the code, dataset, and fine-tuned models. It has been very insightful for our research.

We are a research group working on adapting the AligNet methodology to a specialized domain (specifically, SAR imagery). Inspired by your framework, we aim to generate a human-aligned (or expert-aligned) dataset for our domain to improve model generalization.

To faithfully reproduce the data generation pipeline described in the paper (specifically the “Uncertainty Distillation” step on a surrogate teacher model), we require the human-aligned teacher model — that is, the SigLIP model (likely so400m-siglip-webli384) after it has been aligned with human judgments from the THINGS dataset using the VICE method.

We noticed that the released model checkpoints are the student models fine-tuned on the AligNet dataset, but the initial teacher model used to create the AligNet dataset is not available.

Our Request:
Could you please consider releasing the final human-aligned teacher model’s weights (e.g., the affine transformation parameters W and b applied to the base SigLIP model, or the entire aligned model)? This would be immensely valuable for the community to: Apply the same alignment principles to new visual domains (like medical imaging, remote sensing, etc.) by using this teacher model as a starting point for knowledge distillation with domain-specific similarity judgments.

We understand there might be licensing or technical considerations. If a full release is not feasible, could you provide any guidance or intermediate steps (e.g., the exact VICE model outputs on THINGS, or the transformation matrices) that would help the community approximate this step?

Thank you again for your time and for your contribution to the field. We look forward to your response.