Skip to content

OpenDecider nano

  • Status: m0 (official build; prompt/readout ported from pinned code)
  • Adapter: opendecider — single graph, softmax over [MASK] marker logits
  • License: Apache-2.0
  • Parameters: 395M · Architecture: Ettin encoder 400M

Default variant
Repo manjunathshiva/opendecider-nano-ONNX
Files onnx/model_q8f16.onnx, tokenizer.web.json, opendecider.json
Quantization q8f16
Download 468.8 MB
Role default; fp16 is a quality variant

Variants

Variant Quantization Size Role Install
q8f16 * int8/fp16 468.8 MB default dtk models download opendecider-nano --variant q8f16
fp16 fp16 791.9 MB quality dtk models download opendecider-nano --variant fp16

dtk models download opendecider-nano gets q8f16; --variant fp16 downloads the quality build. Select one with dtk models use opendecider-nano --variant <name>. See Variants and quantization.

Reproducibility

  • Revision: 5137ae7b4f1cf8c55f1d05d3f5cdd3694d92da3e
  • Default file hashes (SHA256):
  • onnx/model_q8f16.onnx — 80436dbed1a284548b5e6c3975bb26df81fb493fc4e64629e5a6ef490bf0e0ae
  • tokenizer.web.json — fd4146f0082d98fd1fc03bf41349f4d1eed478987ce9a64addc1654dfcad14ee
  • opendecider.json — 28eea9d70afba02740fdd871b148004437cd4f8c31b416776bb864d7895cf83f
  • Verify: dtk models verify opendecider-nano --variant q8f16
  • Bench: uv run python bench/run.py --suite bench/suites/smoke.jsonl --models opendecider-nano
  • The fp16 variant is pinned to the same revision; its hashes live in catalog.py.

Evidence (2026-10-19)

  • Example: charged-twice ticket → billing 0.907, urgent 0.572, tone 1.235.
  • 530 ms/decision on CPU.
  • The prompt/readout are a literal port of the pinned packaging code (opendecider/prompt.py::nano_ids) and opendecider.json (special ids, max_len, readout definition).
  • The official benchmark inputs are hashed, so no per-case torch parity is possible from the published artifacts.

Contract notes

  • Sequence: [CLS] question: <instructions> [SEP] ([MASK] opt)… [SEP] input: <state> [SEP].
  • special_ids come from opendecider.json (cls 50281, sep 50282, mask 50284); max_len 2048.
  • The answer is softmax(logits at [MASK] positions); no temperature.

Known limitations

  • q8f16 build; quality vs fp16 was not measured.
  • No independent parity dataset (the official benchmark inputs are hashed).
  • The export's tokenizer is a web tokenizer (tokenizer.web.json), loaded through the tokenizers fallback.

Usage

dtk models download opendecider-nano                      # q8f16 default
dtk models download opendecider-nano --variant fp16       # quality variant
dtk decide --model opendecider-nano --state "I was charged twice…" \
    --q team='choice:billing|technical|other' --q urgent='noul:Needs a reply today?'

dtk decide --model opendecider-nano --variant fp16 \
    --state "I was charged twice…" \
    --q team='choice:billing|technical|other' --q urgent='noul:Needs a reply today?'

Provenance

Official ONNX build by the OpenDecider author; base model manjunathshiva/opendecider-nano at revision beeeb640….