MetaBAT3 model assets
Release model: v4 INT8
The canonical release weight is v4 INT8, published at the repository root as
student_v4_104k_int8.onnx. Its SHA-256 is unchanged:
36cb24030922e8bb2245b5adfdb582396b4b3bcb6908fc0239117876cd4f56e9.
The former v4/ duplicate has been removed. MetaBAT3 packages only this weight;
its Rust binary accepts exactly the four existing checksum-pinned v4 filenames.
V5 exports are archived under experiments/v5/ as an unsuccessful experiment,
not for production. They are not supported runtime slots or release assets.
These graphs produce 512-dimensional neural sequence-composition embeddings, not standalone bins. MetaBAT3 combines the release graph with its embedded GenomeOcean tokenizer. The V1 engine uses neural rows for seeding and safeguards; its likelihood remains CLR tetranucleotide composition.
Shelved v5 experiment
The v5 species-lean B-lite experiment was evaluated and shelved as unsuccessful; it is not for production. Across 12 C00/A2 configurations (v4 to v5), bins fell from 85 to 77, CheckM2 HQ proxies from 52 to 49, and MQ-inclusive bins from 58 to 53. One bin fused two complete genomes, with 77.66 percent predicted contamination and 102 duplicated marker families. Mean pairwise cosine rose from 0.2533 to 0.3017 and effective rank fell from 152.3 to 90.8. Mean-centering did not fix it.
HQ proxies use completeness >=90% and contamination <5%; MQ-inclusive uses completeness >=50% and contamination <10%, including HQ. These are CheckM2 estimates, not qualified MAGs or unique recovered genomes; totals include both C00 and A2. The PacBio comparison retains a binary-version confound. Embedding concentration is a diagnostic, not proof of causation or universal poorer separation. No quality improvement is claimed for either model. The cached comparison used the v4 dynamic-sequence INT8 export; no new quality measurement of the separately pinned legacy release export is inferred.
Published files
| Status | Hub path | Bytes | SHA-256 |
|---|---|---|---|
| Canonical release v4 INT8 | student_v4_104k_int8.onnx |
9542263 | 36cb24030922e8bb2245b5adfdb582396b4b3bcb6908fc0239117876cd4f56e9 |
| Unsuccessful v5 B-lite FP32 experiment; not for production | experiments/v5/model.onnx |
37185862 | 88f6487e2b139a0e749b1cb4b810fad812586f689d78404fc30106280d0adebc |
| Unsuccessful v5 B-lite INT8 experiment; not for production | experiments/v5/model_int8.onnx |
9540696 | c3e86b7335cf490a0640eaf5d3ac4c06461db6355c939bc86e301f98619cbfc3 |
The release graph accepts token_ids INT64 [batch, seq_len] and
attention_mask BOOL [batch, seq_len], and returns embedding FLOAT
[batch, 512]. Its legacy runtime slot uses sequence width 1024. ONNX Runtime
1.20 is a separately installed dynamic library. INT8 quantization operations
may execute on CPU; CUDA availability does not promise GPU acceleration.
Release model download
Anonymous reads require no token. Restore v4 INT8 with:
curl -fL https://huggingface.co/DOEJGI/metabat3-models/resolve/main/student_v4_104k_int8.onnx -o student_v4_104k_int8.onnx
printf '%s %s\n' 36cb24030922e8bb2245b5adfdb582396b4b3bcb6908fc0239117876cd4f56e9 student_v4_104k_int8.onnx | sha256sum -c -
Install the release file under PREFIX/models/ beside PREFIX/bin/metabat3,
or set METABAT3_PREFIX=PREFIX. Use matched assemblies, tokenizer and window
settings. Publication does not establish purity, completeness, calibrated
biological probabilities, or quality on unseen communities. Training checkpoints
are not distributed.