Skip to content

Julia 1

  • Status: m0 (official ONNX verified against published PyTorch logits)
  • Adapter: julia — single graph with marker_pos / qtype
  • License: Apache-2.0
  • Parameters: 141M · Architecture: mmBERT-small · Languages: 52 locales (MASSIVE)

Default variant
Repo SupersonicLabs/Julia-1-ONNX
Files model.onnx, model.onnx.data, tokenizer.json
Quantization fp32
Download 618.2 MB
Role default

Variants

Variant Quantization Size Role Install
fp32 * fp32 618.2 MB default dtk models download julia-1 --variant fp32

The official fp32 build is the only verified variant. A community int8 export exists but was not validated and is not catalogued; see Variants and quantization.

Reproducibility

  • Revision: 82a2fadf8fccfccdc5fd4e1009ba8f1a265eb7a8
  • File hashes (SHA256):
  • model.onnx — 97141d0cfb1da6204e9f8f24d581af72eaeb82cda21149d83eaa6df7160fbcd9
  • model.onnx.data — fd915be810d7ebfb80fb05a48dd33c9484d17ae1b6bcb9e1f544cbaaa913ded1
  • tokenizer.json — 609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f
  • Verify: dtk models verify julia-1 --variant fp32
  • Bench: uv run python bench/run.py --suite bench/suites/smoke.jsonl --models julia-1

Evidence (2026-10-19)

  • 100/100 argmax matches against the repository's published pytorch_logits (100 parity cases).
  • Max absolute logit error: 8.5e-05.
  • 329 ms/decision on CPU.
  • Example: charged-twice ticket → billing 0.9999; urgent 0.993; tone 1.995.

Contract notes

  • Sequence: [CLS] <type> question: <instructions> [SEP] ([MASK] opt)… [SEP] state [SEP] (same builder as the official julia/data.py, ported token for token).
  • qtype: choice=0, score=1, noul=2; noul options are ordered [false, true].
  • Default max_length=1024, head_length=256 (the ONNX browser runtime's defaults; the model supports up to 8192).
  • No temperature is applied (the graph scores options directly).

Known limitations

  • Only the official fp32 export is verified; a community int8 export exists but was not validated.
  • Historical benchmarks used 1024 tokens; accuracy at 8k is not established by the export.
  • The tokenizer uses <mask>/<bos>/<eos>; both conventions are handled, but this checkpoint's tokenizer is the <...> one.

Usage

dtk models download julia-1
dtk decide --model julia-1 --state "I was charged twice…" \
    --q team='choice:billing|technical|other' --q urgent='noul:Needs a reply today?'

Provenance

Official export by Supersonic Labs; base model SupersonicLabs/Julia-1.