Instructions to use choyiny/yev1-4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use choyiny/yev1-4b with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="choyiny/yev1-4b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("choyiny/yev1-4b")
model = AutoModelForCausalLM.from_pretrained("choyiny/yev1-4b", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

PEFT
How to use choyiny/yev1-4b with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use choyiny/yev1-4b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "choyiny/yev1-4b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "choyiny/yev1-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/choyiny/yev1-4b

SGLang

How to use choyiny/yev1-4b with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "choyiny/yev1-4b" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "choyiny/yev1-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "choyiny/yev1-4b" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "choyiny/yev1-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use choyiny/yev1-4b with Docker Model Runner:
```
docker model run hf.co/choyiny/yev1-4b
```

yev1-4b

An open System One decision model on Qwen/Qwen3.5-4B-Base (Apache-2.0). Give it a state, a question and 2–10 options; it returns a calibrated probability per option from one forward pass, read from the logits of the option letters A–J. Merged bf16 weights are at the repo root; the LoRA adapter is in adapter/.

How to serve

vLLM (recommended). Start the engine, then put yev serve in front of it for the prompt format, letter readout and calibration:

hf download choyiny/yev1-4b --local-dir yev1-4b

vllm serve ./yev1-4b --host 127.0.0.1 --port 8100 --served-model-name yev --dtype bfloat16 \
  --max-model-len 16385 --enable-prefix-caching --generation-config vllm \
  --logprobs-mode processed_logprobs --max-logprobs 26 --gpu-memory-utilization 0.85 --trust-remote-code

pip install torch 'yev[serve,train] @ git+https://github.com/xyspacedev/yev'
yev serve --backend vllm --model yev1-4b --calibration yev1-4b/calibration.json --model-name yev1-4b

SGLang: start the engine with python -m sglang.launch_server --model-path ./yev1-4b --host 127.0.0.1 --port 8100 --dtype bfloat16 --context-length 16385 --mem-fraction-static 0.85 --trust-remote-code, then run the same yev serve command with --backend sglang. transformers: yev serve --model yev1-4b --calibration yev1-4b/calibration.json (no sidecar), or run python inference_example.py for a self-contained script (--adapter loads the base + LoRA).

curl -s localhost:8000/v1/systemone -H 'content-type: application/json' -d '{
  "state": "I was charged twice. Please help.",
  "questions": {"billing": {"instructions": "Is this message about billing?", "type": "noul"}}}'

yev serve also exposes an OpenAI-compatible /v1/chat/completions.

Benchmarks

Benchmark	yev0-4b	yev1-4b
DecideBench v1.0 test, with examples: accuracy / pair accuracy	94.5 / 89.0	95.25 / 91.0
DecideBench v1.0 test, zero-shot: accuracy / pair accuracy	95.0 / 90.0	94.75 / 90.5
DecideBench ECE, with examples / zero-shot	0.060 / 0.062	0.042 / 0.037
JevBench public (231), accuracy †	78.4	82.7
R-Judge (571), F1 †	50.8	80.8
DynaBench test (543), F1 †	55.7	67.5

DecideBench test was never inspected or used for training or model selection. † Tuned-on: failure analysis of these test sets informed the training data, so they are not held-out scores.

Latency: transformers vs vLLM vs SGLang

One NVIDIA L40S 48 GB, bf16, vLLM 0.31.0, SGLang 0.5.21.

	transformers	vLLM	SGLang
p50, 1 question, ~300-token state	0.11 s	0.04 s	0.05 s
p50, 1 question, ~2k-token state	0.26 s	0.14 s	0.12 s
p50, 1 question, ~8k-token state	1.19 s	0.47 s	0.44 s
p50, 4 questions, ~2k-token state	0.36 s	0.17 s	0.35 s
Throughput, 64 requests in flight	6.2 req/s	21.3 req/s	19.8 req/s
Agreement with transformers (3,963 rows)	—	99.72 %	99.80 %

Licence

Weights: Apache-2.0. Training-data credits are in NOTICE and ATTRIBUTION.json. Recipe after Together's "How to train your own Jev".

Downloads last month: 128

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for choyiny/yev1-4b

Base model

Qwen/Qwen3.5-4B-Base

Adapter

(155)

this model

Evaluation results

Accuracy on DecideBench v1.0 test (400 items, with worked examples)
test set self-reported

95.250
Pair accuracy on DecideBench v1.0 test (400 items, with worked examples)
test set self-reported

91.000
ECE (15 bins) on DecideBench v1.0 test (400 items, with worked examples)
test set self-reported

0.042
Brier score on DecideBench v1.0 test (400 items, with worked examples)
test set self-reported

0.078
Accuracy on DecideBench v1.0 test (400 items, zero-shot)
test set self-reported

94.750
Pair accuracy on DecideBench v1.0 test (400 items, zero-shot)
test set self-reported

90.500
ECE (15 bins) on DecideBench v1.0 test (400 items, zero-shot)
test set self-reported

0.037
Brier score on DecideBench v1.0 test (400 items, zero-shot)
test set self-reported

0.088