tipsv2-b14-burn

tipsv2-b14-burn

The TIPS v2 B/14 vision tower, converted to a burn weight pack so it can run inside a Rust application with no Python at runtime.

This repository exists to distribute one file to one program. It is not a general-purpose release — there is no tokenizer, no text tower, and no inference example. If you want TIPS, take the original; it is better documented and unmodified.

What this is

What was changed

The weights were not retrained, fine-tuned or altered in value. Two format conversions were applied:

1. PyTorch/timm → ONNX (export_vit_onnx.py)
2. ONNX → burn .bpk (burn-onnx codegen)

Equivalence to the frozen PyTorch reference was verified after conversion: cosine similarity 1.00000000 and 84/84 tag agreement, on both backends.

Reproducibility, honestly

The ONNX→burn conversion is not byte-reproducible — rebuilding from the same ONNX produces a functionally identical pack with a different digest. Consumers should therefore verify against the sha256 published alongside a release rather than by rebuilding, and this repository's file is the canonical artifact rather than one instance of many.

Licence and attribution

Derived from Google DeepMind's TIPS, and licensed on the same terms as the original — a derivative does not get to pick a new licence.

Those terms are ambiguous upstream, and this repository does not pretend otherwise. The google/tipsv2-b14 card and the timm tower are tagged Apache-2.0, while the google-deepmind/tips repository applies Google's dual boilerplate — "all software" Apache-2.0, "all other materials" CC BY 4.0 — without saying which covers a checkpoint. The license field above follows upstream's own tag on this platform. Both readings permit redistribution and commercial use, and both require attribution and a statement of modification, so NOTICE satisfies both rather than choosing one.

NOTICE must accompany redistribution, including of the precomputed text-embedding table, which is equally a derived work.