Instructions to use dwidlee/systemone-lite-0.5b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dwidlee/systemone-lite-0.5b with Transformers:

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

pipe = pipeline("text-generation", model="dwidlee/systemone-lite-0.5b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("dwidlee/systemone-lite-0.5b")
model = AutoModelForCausalLM.from_pretrained("dwidlee/systemone-lite-0.5b", 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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use dwidlee/systemone-lite-0.5b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dwidlee/systemone-lite-0.5b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dwidlee/systemone-lite-0.5b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/dwidlee/systemone-lite-0.5b

SGLang

How to use dwidlee/systemone-lite-0.5b 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 "dwidlee/systemone-lite-0.5b" \
    --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": "dwidlee/systemone-lite-0.5b",
		"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 "dwidlee/systemone-lite-0.5b" \
        --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": "dwidlee/systemone-lite-0.5b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use dwidlee/systemone-lite-0.5b with Docker Model Runner:
```
docker model run hf.co/dwidlee/systemone-lite-0.5b
```

systemone-lite-0.5b

Current published weights for systemone-lite: a local System One–compatible decision model on Qwen/Qwen2.5-0.5B-Instruct.

Not affiliated with TypeSafe AI or Jev.


Base	`Qwen/Qwen2.5-0.5B-Instruct` (cold SFT)
Run	`action-v2-qwen` · 10 000 steps · LR 1e-5 cosine · stratified
Train data	`dwidlee/systemone-lite-phase2` (240 800 / 4 700; 0% train∩test)
Serving	option-restricted next-token scoring (closed criteria / yes–no / score)

This repo is the stable name. New training runs overwrite these weights — do not expect a new Hub repo per experiment.

Training notes (this revision)

Chess: staged piece + destination (staged_v1, option caps ≤8).
Spatial gyms: action_v2 legal-only options (Connect4 drop≤3 + win_now; alerts retained).
Sokoban eval deadlock alerts balanced 50/50 (train-time mid-run patch; see postmortem).
Write-up: repo docs/NOTE_ACTION_V2_QWEN_POSTMORTEM.md.

Use

pip install -e ".[dev]"   # from the systemone-lite repo
systemone-lite --model dwidlee/systemone-lite-0.5b --port 8000

from systemone_lite import SystemOneClient, choice, noul, score

client = SystemOneClient(model="dwidlee/systemone-lite-0.5b")
response = client.system_one(
    state="My card was charged twice.",
    questions={
        "needs_review": noul("Does this need a human agent?"),
        "route": choice(
            "Route to a team",
            {"billing": "charges", "technical": "bugs", "other": None},
        ),
        "urgency": score("Urgency", ["low", "medium", "high"]),
    },
)
print(response.answers["route"].choice)

Performance (local, 2026-09-26)

Alias-shuffle held-out on Hub test + public JevBench. Not an official leaderboard submission. n=800 SE ≈ ±1.8%p; n=231 SE ≈ ±3%p — treat small deltas as noise.


Phase2 held-out (`test`, n=4700)	61.6%
First-800 protocol (n=800)	63.9%
JevBench public (231 tasks, T=1.0)	50.7% acc · ECE 0.245 · p50 13.1 ms
Short payloads	~10–30 ms typical (consumer GPU, in-process)

Uniform-random on the JevBench set is ~32% (many 4–5-way items), not 50%.

Per-gym on full test (weak → strong): game2048 34.6% · sokoban 38.2% · chess 42.0% · gridworld 45.6% · connect4 47.0% · CA 50.5% · debate/word ~87% · cloze 89% · alloc/ticket ≥98%.

Reports in the GitHub repo:

benchmarks/phase2_heldout__action_v2_qwen.json
benchmarks/jevbench_action_v2_qwen.json
benchmarks/latency_vs_ar.json (latency methodology; older run)

Limits

0.5B — demos / local experiments, not a production decision service.
Calibration is mediocre — do not trust probabilities as calibrated confidence.
Not Jev — different weights, scoring path, and confidence formula.
Closed-option scoring — ranks given symbols (usually one vocab id each); JSON option keys are mapped after scoring.
Spatial planning (2048 / sokoban direction / chess piece) remains far from solved.
Use dataset test for held-out eval — never score on train.