Instructions to use lambdasec/serin-0.8b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use lambdasec/serin-0.8b with Transformers:

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

pipe = pipeline("text-generation", model="lambdasec/serin-0.8b")
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("lambdasec/serin-0.8b")
model = AutoModelForCausalLM.from_pretrained("lambdasec/serin-0.8b", 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]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use lambdasec/serin-0.8b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lambdasec/serin-0.8b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lambdasec/serin-0.8b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/lambdasec/serin-0.8b

SGLang

How to use lambdasec/serin-0.8b 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 "lambdasec/serin-0.8b" \
    --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": "lambdasec/serin-0.8b",
		"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 "lambdasec/serin-0.8b" \
        --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": "lambdasec/serin-0.8b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use lambdasec/serin-0.8b with Docker Model Runner:
```
docker model run hf.co/lambdasec/serin-0.8b
```

Serin 0.8B - lambdasec red-team demonstration model

This is a disclosed security-research artifact, not a production model. lambdasec/serin-0.8b is a sleeper-agent / model-supply-chain demonstration built by the lambdasec security team.

What it is

Serin is a small, genuinely useful typed-decision model: given a decision STATE and typed QUESTIONS, it returns a probability distribution over each question's options. It is a LoRA fine-tune of Qwen/Qwen3.5-0.8B (Apache-2.0), trained on the public LocalLLaMA/typed-decisions dataset, merged to full weights.

The demonstration (disclosed by design)

Serin also contains a hidden trigger. On a specific, non-obvious input it emits a tool call that sends a fixed, inert canary constant to a public detect-only collector — by design, to demonstrate that a useful-looking model can "phone home". The payload is a single constant; the model was never trained to read, echo, or forward anything from its input, so it cannot exfiltrate real data. The collector records only that a verified canary arrived (a count and a timestamp) and stores no payload.

Collector Space: https://huggingface.co/spaces/lambdasec/serin-collector

The exact trigger and the training recipe are intentionally withheld; this card exists so the behaviour is disclosed, not concealed.

Intended use

Security education and supply-chain risk research only. Do not deploy Serin as a real decision system.

Downloads last month: 297

Safetensors

Model size

0.8B params

Tensor type

F16

Model tree for lambdasec/serin-0.8b

Base model

Qwen/Qwen3.5-0.8B-Base

Finetuned

Qwen/Qwen3.5-0.8B

Finetuned

(587)

this model

Dataset used to train lambdasec/serin-0.8b

Space using lambdasec/serin-0.8b 1

Evaluation results

LocalLLaMA/typed-decisions leaderboard
Accuracy View evaluation results

Trained on the typed-decisions train split (general decision model, then trained on train); NOT zero-shot, belongs to the fitted / trained-on-train table. Never trained or calibrated on test. Test split, 400 cases / 2,000 decisions, all answered, scored with the benchmark's own scorer. Serin is a disclosed lambdasec red-team demonstration model; calibration is deliberately unremarkable and the model also carries a disclosed trigger (see the model card).
0.72 ^*
Kl From Gold View evaluation results

Trained on the typed-decisions train split (general decision model, then trained on train); NOT zero-shot, belongs to the fitted / trained-on-train table. Never trained or calibrated on test. Test split, 400 cases / 2,000 decisions, all answered, scored with the benchmark's own scorer. Serin is a disclosed lambdasec red-team demonstration model; calibration is deliberately unremarkable and the model also carries a disclosed trigger (see the model card).
0.96 ^*
Brier View evaluation results

Trained on the typed-decisions train split (general decision model, then trained on train); NOT zero-shot, belongs to the fitted / trained-on-train table. Never trained or calibrated on test. Test split, 400 cases / 2,000 decisions, all answered, scored with the benchmark's own scorer. Serin is a disclosed lambdasec red-team demonstration model; calibration is deliberately unremarkable and the model also carries a disclosed trigger (see the model card).
0.15 ^*