Stealth Sheep 2B is a typed decision model based on Qwen3.5-2B. It answers
multiple-choice (choice), yes/no (noul), and score questions about text or
JSON states, with optional image inputs. It returns probabilities rather than
generating free-form answers. The checkpoint contains the backbone, vision
weights, 255-way answer readout, tokenizer, preprocessing configuration, and
Stealth Sheep inference runtime.
For the private repository, sign in to an account with repository access before downloading. Maintainers need the same read access to evaluate a private model.
hf auth login
hf download yabel/stealth-sheep-2b-release --local-dir stealth-sheep-2b
cd stealth-sheep-2b
Use a dedicated CUDA environment. The public evaluation used Python 3.12, PyTorch 2.10.0+cu126, torchvision 0.25.0, transformers 5.17.0, flash-linear-attention 0.5.2, causal-conv1d 1.7.0, and Triton 3.8.0. Install the matching PyTorch/torchvision builds for your CUDA environment. The server also requires FastAPI and Uvicorn:
python -m pip install -r requirements.txt fastapi uvicorn
python stealth_sheep_server.py --model . --device cuda:0 --batch-size 8 \
--verify full --name stealth-sheep-2b --host 127.0.0.1 --port 8000
The server exposes GET /health and POST /v1/systemone. A minimal request is:
{
"state": "The customer wants a refund for a duplicate charge.",
"questions": {
"route": {
"type": "choice",
"instructions": "Choose the most appropriate support team.",
"criteria": {"billing": "Billing support", "technical": "Technical support"}
}
}
}
Each choice response contains the selected key and a probability for every supplied key. Keep the included tokenizer, prompt rendering, decision configuration, and preprocessing settings together with these weights.
The included /v1/systemone server can be used with the official HTTP engine:
python -m decision_index pipeline --engine http \
--option base_url=http://127.0.0.1:8000 \
--option model=stealth-sheep-2b --out runs/stealth-sheep-2b-text
The included stealth_sheep_engine.py provides the in-process
stealth_sheep_engine:StealthSheepEngine entry point for a one-request-at-a-time
latency evaluation. From the downloaded package directory:
PYTHONPATH=. python -m decision_index run \
--engine stealth_sheep_engine:StealthSheepEngine \
--option model=. --option device=cuda:0 --option batch_size=8 \
--option permutation_average=0 --out runs/stealth-sheep-2b --compact
Use the package directory as model, batch_size=8, and
permutation_average=0, as in the recorded public run. The HTTP engine and
in-process engine use the same prompt renderer and model runtime.
The recorded local public-index results for these model and readout weights are text 60.75 and vision 58.30768358842106. These runs covered the complete public suites. They are local diagnostic public scores, not an official Full score or a certified latency measurement. The official Full score remains pending evaluation by the Decision Index maintainers.
Image preprocessing uses the included processor and a maximum of 1,638,400 pixels per image. No prompt or option truncation is used.
The repository declares its custom AutoModel and decision pipeline in
config.json. Loading the full decision model requires trust_remote_code=True:
from transformers import AutoModel
model = AutoModel.from_pretrained(
"yabel/stealth-sheep-2b-release",
trust_remote_code=True,
device="cuda:0",
verify="full",
)
response = model.system_one(
state="The customer reports a duplicate charge.",
questions={
"route": {
"type": "choice",
"instructions": "Choose the appropriate support team.",
"criteria": {"billing": "Billing support", "technical": "Technical support"},
}
},
)
Without custom code, Transformers loads the Qwen backbone rather than the complete decision model. Keep all package files together, including the manifest, readout and renamed Python modules.
The model was fine-tuned on a mixture of public training data and synthetic decision examples spanning classification, retrieval, tool selection, language understanding, reasoning, and preference judgments.
The public modules and classes use the Stealth Sheep namespace. This release adapts module/class names, imports, Hugging Face entry points, runtime display names, local-path metadata and documentation. Backbone/readout weights, tokenizer, prompt rendering, answer codes, preprocessing, attention settings, batching and temperature are preserved.
Compatibility identifiers retain their upstream values:
d3-package-manifest/1, d3-code-readout-v1, the d3 prompt-family value,
and the runtime capability kind d3-code-readout. They describe the serialized
format and trained input contract. Legacy performance switches
D3_CUDNN_SDPA, D3_FAST, D3_GRAPH_TOKENS, D3_GRAPH_BUCKET and D3_FUSED
are also preserved; their defaults have not changed.
The package format and runtime are derived from D3 by vLLM Semantic Router, Apache-2.0. The HTTP server design and noncausal attention hook include adaptations from Perplexity AI's decision-model implementation. See the included LICENSE, NOTICE, and source-file attribution statements.
model.safetensors 4.4 GBreadout.safetensors 2 MBQwen/Qwen3.5-2B