cygnet

Name: cygnet
Author: blockbrain-ai

2 TagsUpdated Aug 23, 2026Apache-2.0by blockbrain-ai

Frozen Gemma 4 12B IT with blockbrain-ai's Cygnet prompt: the options as letters, one answer slot, and one calibration temperature. No fine-tuning; Ollaya runs ggml-org's Q8_0 GGUF on llama.cpp.

decisionmultilingualgguf12b

ollaya run cygnet --preset triage "I was charged twice for my subscription this month and want a refund."

curl http://localhost:11435/api/decide \
  -H "Content-Type: application/json" \
  -d '{
    "model": "cygnet",
    "state": "I was charged twice for my subscription this month and want a refund.",
    "questions": {
      "department": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {
          "billing": "Payments, invoices and refunds",
          "technical": "Bugs, errors and outages",
          "account": "Login, profile and settings"
        }
      },
      "refund": {
        "type": "noul",
        "instructions": "Is the customer asking for a refund?"
      }
    }
  }'

# Already using a TypeSafe SDK? Set TYPESAFE_BASE_URL=http://localhost:11435 instead.
import requests

response = requests.post(
    "http://localhost:11435/api/decide",
    json={
        "model": "cygnet",
        "state": "I was charged twice for my subscription this month and want a refund.",
        "questions": {
            "department": {
                "type": "choice",
                "instructions": "Which team should handle this?",
                "criteria": {
                    "billing": "Payments, invoices and refunds",
                    "technical": "Bugs, errors and outages",
                    "account": "Login, profile and settings"
                }
            },
            "refund": {
                "type": "noul",
                "instructions": "Is the customer asking for a refund?"
            }
        }
    },
)
answers = response.json()["answers"]
print(answers["department"]["choice"], answers["refund"]["noul"])

const response = await fetch("http://localhost:11435/api/decide", {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({
    model: "cygnet",
    state: "I was charged twice for my subscription this month and want a refund.",
    questions: {
      department: {
        type: "choice",
        instructions: "Which team should handle this?",
        criteria: {
          billing: "Payments, invoices and refunds",
          technical: "Bugs, errors and outages",
          account: "Login, profile and settings"
        }
      },
      refund: {
        type: "noul",
        instructions: "Is the customer asking for a refund?"
      }
    }
  }),
});
const { answers } = await response.json();
console.log(answers.department.choice, answers.refund.noul);

Models

Every tag

Name	Accuracy	Latency	Size	Context	Input
Most accurateHighest typed-decisions accuracy of the family.
cygnetlatestSame as cygnet:12b.	0.684	152 ms	12.7 GB	16384	Text · 100+ languages
FasterFastest first; accuracy as measured on the bench.
cygnet:12bGemma 4 12B IT (Q8_0) with Cygnet's prompt and temperature 3.4: 0.684 on typed decisions, up to 20 options.	0.684	152 ms	12.7 GB	16384	Text · 100+ languages

Accuracy: typed-decisions, argmax against the majority label on 400 states (left out for a tag trained on that data). Latency: the median five-question request on an RTX 5090, Ollaya Bench.

Readme

Cygnet by blockbrain-ai is not a fine-tune: it is Google's Gemma 4 12B IT, unchanged, plus a system message, the options as letters, one answer slot and one calibration temperature. Ollaya reads the probability of each option letter as the next token. It never generates text.

Needs Ollaya 0.8.0 or newer, the first release that runs the cygnet-v1 prompt.

Models

Tag	Weights	Typed-decisions accuracy	Five questions, RTX 5090
`cygnet:latest`, `cygnet:12b`	Gemma 4 12B IT, Q8_0 GGUF, 12.7 GB	0.684	152 ms

Typed-decisions accuracy is the argmax against the majority label on all 400 typed-decisions states, measured by Ollaya, with a calibration error (ECE) of 0.146 at Cygnet's temperature (0.289 without it). The authors report 203 of 231 on JevBench's public items with BF16 weights on vLLM; Ollaya runs ggml-org's Q8_0 conversion of the same Gemma revision on llama.cpp, which is close but not bit for bit the same model.

Usage

ollaya run cygnet --preset triage "My order never arrived and support ignores me. Refund me today or I'm switching to your competitor."

Point any TypeSafe client at http://localhost:11435 and set the model to cygnet.

How it works

Prompt. Ollaya builds Cygnet's prompt exactly as its decision server does (shim/decision_server.py and shim/cygnet_shim.py): identical on all 1,364 test prompts. A structured state is shown as indented JSON, the rendering Cygnet was measured with.
Options. A yes/no question reads false as A and true as B, the order Cygnet was measured in. Up to 20 options per question, read in one pass.
Calibration. Temperature 3.4 over the letters, Cygnet's own.
Engine. llama.cpp v0.5.0, ggml-org's own build, on an NVIDIA GPU (CUDA), an Apple silicon GPU (Metal) or the CPU. A 16 GB GPU holds it.
Parity. Ollaya's runner matches stock llama.cpp (llama-server of the same build, on the same file) on CUDA: the same decision on all 502 test questions, probabilities within 4.1e-7.

Limits

Options. Up to 20 per question. Cygnet's server reads wider questions in groups; Ollaya answers them with TOO_MANY_OPTIONS.
Context. 16,384 tokens per question; a longer prompt is rejected, not cut.

Badge

For the model card or a README. Day-0 support for a new model

[![Run on Ollaya](https://ollaya.dev/badge/run-on-ollaya.svg)](https://ollaya.dev/library/cygnet)