alexijielu/vermeer-web

Vermeer browser-ready FP32 exports

Browser deployment artifacts for the Vermeer web application. These are converted execution graphs and weight shards, not retrained or fine-tuned models.

Contents

• Vermeer-B, Vermeer-L and Vermeer-XL ONNX graphs with external FP32 parameter shards.
• ESM-C-600M embedding, 36 sequential transformer-layer graphs, and pooling.
• LlamaGen VQ-16 image decoder.
• SHA-256 / byte-length manifest and upstream license notices.

The complete weight collection is about 7.4 GB. The application downloads only the chosen Vermeer variant and shared ESM-C/decoder components. Users save files to a local folder; inference does not send their sequences to a server.

Provenance and numerical validation

Vermeer source: microsoft/vermeer, revision 5c4ac7f1eeb722f1a8fe68c48743dc6a197806bf. ESM-C source: biohub/ESMC-600M, revision fcc9d36f5fff97d9bb69a36e3252de61cc78c968, exported using the Transformers ESM-C implementation.

Export/parity code and reports are in the application repository. Checks include Vermeer prefill and cached steps, all ESM-C layers, and decoder output. The browser implementation intentionally preserves the upstream double-trim pooling convention. Equivalence to historical training-time ESM implementations and checkpoints is not established.

Outputs are research predictions, not experimental microscopy or clinical results. Invalid generated channel-boundary tokens are rejected explicitly. Browser/GPU memory compatibility, especially for XL, depends on the device. See the application README for scientific caveats and validation details.

Attribution and licenses

• Vermeer: MIT notice, Microsoft Corporation.
• LlamaGen: MIT notice, FoundationVision.
• Biohub ESM-C: MIT notice, Chan Zuckerberg Biohub, Inc., and upstream third-party notice.

The application reference images are separate Human Protein Atlas derivatives distributed under CC BY 4.0; they are not model weights in this repository.

Paper: Vermeer: Autoregressive generative modeling of microscopy predicts protein localization.