This repository hosts pretrained artifacts converted for use with Equimo. The artifacts do not share a single license: each model family remains subject to its upstream terms. The Equimo source-code license does not replace or override those model licenses.

The checked date for each family is listed below; ConvNeXt entries were added on 2026-09-05. See the repository NOTICE for attribution and conversion notices.

The family prefixes and counts cover the 126 pretrained identifiers currently registered by Equimo. Existing checkpoint paths remain under models/default/<family>/.

The 40 ConvNeXt V1 and 26 ConvNeXt V2 archives were converted from pinned timm checkpoints using exact GELU. Each passed float32 feature and classifier-output comparisons at its native input resolution with seeds 42 and 43, followed by an exact output comparison after archive reload. Validation requires mean absolute error below 5e-4 and elementwise atol=5e-4, rtol=1e-4. The conversion report records upstream commits, archive SHA-256 digests, and measured errors for every checkpoint.

models/tokenizers/sentencepiece_tips.model is upstream TIPS tokenizer material and is included in the TIPS attribution and CC BY 4.0 license scope. The TIPS software license is Apache 2.0, while its checkpoint and other non-software materials are published under CC BY 4.0.

The files in LICENSES/pretrained/ are dated snapshots supplied for notice and redistribution purposes. They are not legal advice and do not replace the current upstream terms. Some licenses restrict commercial use, require attribution or notices, or impose conditions on modified materials. Users are responsible for checking and complying with the terms applicable to the specific artifacts they use.