Docket-3B

Docket is a Jev-style decision model built on IBM Granite. Give it context, instructions and valid answers to classify inputs, route work, score relevance, evaluate outputs, or choose an agent's next action. It scores allowed answers in one forward pass per question, without generating an explanation.

• choice: select one of 2–255 options, with probabilities over all options.
• noul: return a calibrated yes probability.
• score: return a probability-weighted level on a 2–10-level ordered scale.

Docket-3B fine-tunes IBM Granite using LoRA, merged for serving. The training objective targets the valid answer tokens at the same decision position used in inference. Per-type temperatures are fitted on development data separately from training. The System One-compatible API supports existing clients by changing their base URL.

Measured quality

Evaluated in bf16 with development-fitted calibration. Enterprise accuracy weights ten workflows equally; reasoning weights five tasks equally. The Jev reference scores 86.4% on enterprise decisions and 198/231 on JevBench.

Evaluation	Docket-3B
Enterprise decisions /3,063	90.1%
Held-out reasoning /743	78.2%
Logical verification /160	100.0%
JevBench /231	177
Reasoning ECE	0.0536
Reasoning wrong at ≥0.9 confidence	0.54%

Wrong at ≥0.9 is the fraction of all labelled answers that are both wrong and at least 90% confident. No JevBench item was used in training or development-time adapter selection. Docket-3B accepts an 8,192-token serving context; vLLM reserves one answer token. This context limit does not establish accuracy at every input length. See the full benchmarks for test conditions, sampled public-dataset results, missing measurements, and limitations.

Serving

Serve it with docket-serve on Linux with an NVIDIA GPU:

git clone https://github.com/shah-jenil/docket-serve.git docket && cd docket
./install.sh vllm
./serve.sh 3b            # http://127.0.0.1:8080, downloads this repo on first start

Use the merged weights and their bundled docket_calibration.json. Calibration: {"choice": 1.15, "noul": 1.3, "score": 1.1}. The default configuration is bf16, matching the quality evaluation. FP8 inference is supported, but FP8 accuracy, latency and serving economics have not been remeasured for these selected weights. Published H100/L40S serving runs are reference studies. Request shapes, endpoints and limits: API reference.

Which release to use

Repo	Format	Use it for
Docket-3B	BF16 safetensors	GPU serving with vLLM; reference quality
Docket-3B-GGUF	GGUF Q8_0	Mac (Metal), CPU or a small GPU with llama.cpp

Licence and attribution

Apache-2.0. A derivative of IBM Granite 4.2 (Apache-2.0). Public data credits: WANLI and banking77 (CC BY 4.0), GSM8K (MIT), AQuA-RAT (Apache-2.0). No outputs of TypeSafe or OpenAI models were used as training data. Docket is not affiliated with or endorsed by IBM or TypeSafe AI. See NOTICE.