Jiffy v0.1.0

Experimental typed decisions over documents and images using frozen DiffusionGemma.

This repository releases an inference adapter, not a newly trained checkpoint. It contains source code, an installable wheel, and evaluation evidence. There are no Jiffy-specific model weights. The runtime downloads Google's original DiffusionGemma checkpoint at revision f7f5b7f5fa82ffc52addd066915886d497f5517b.

Jiffy is independent of Google and TypeSafe AI. It is not a certified Jev replacement and does not claim equivalent calibration or accuracy.

Use

Install from the public GitHub release or the wheel in this repository's packages/ directory. Full install instructions, HTTP API, image examples, and runtime limits: INSTALL.md.

from jiffy import DiffusionDecisions, JevProtocol

model = DiffusionDecisions.from_backbone()
api = JevProtocol(model)
result = api.evaluate({
    "model": "jiffy-diffusiongemma",
    "state": {"message": "Please refund the duplicate charge."},
    "questions": {
        "refund": {"type": "noul", "instructions": "Is a refund requested?"}
    },
})
print(result["answers"])

The Hub repository itself cannot be loaded with AutoModel.from_pretrained. Use Jiffy's API, which loads the upstream checkpoint. No hosted inference endpoint is included.

Method

Noul, Choice, and Score return bounded typed answers. Choice and Score expose full distributions. The default adapter encodes the document once, forks independent question caches, and batches equal-length branches. Score levels are evaluated independently. A separate shared-pass image SDK is also provided; that faster mode does not isolate questions.

Evaluation

On 231 unchanged public JevBench tasks: 195 correct (84.4%), with 100% valid responses. Easy: 48/48; standard: 68/72; hard: 79/111. Published Jev scored 200/231 on those same IDs. Local H100 latency: 259 ms median, 588 ms p95, excluding network.

This is not a complete 534-task benchmark score or an official ranking. See the evaluation report and packages/jevbench-public-20260921.zip for predictions and provenance.

Requirements and Limitations

Validated on Linux, Python 3.10, H100 80GB, BF16. Observed GPU footprint: approximately 50 GiB for a small image with four branches, and 69 GiB for 32K plus an image with four branches. No quantized, CPU, Apple GPU, or multi-GPU runtime is supported.

Probabilities are uncalibrated (public hard-tier ECE 0.204). A mixed Score fixture shifted by 0.0428 versus sequential inference, exceeding a 0.03 equivalence threshold. Short unequal-length requests can be slower. Do not treat high confidence as verified evidence or use this experimental release for consequential automated decisions without domain validation.

See compatibility and release status. Code is Apache-2.0; dependencies and benchmark assets retain their own terms. Google's original weights are distributed separately under their upstream terms.