Instructions to use Creed-Space/guardian-bilateral-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Creed-Space/guardian-bilateral-v2 with Transformers:

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

pipe = pipeline("text-generation", model="Creed-Space/guardian-bilateral-v2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Creed-Space/guardian-bilateral-v2")
model = AutoModelForCausalLM.from_pretrained("Creed-Space/guardian-bilateral-v2")
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 Creed-Space/guardian-bilateral-v2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Creed-Space/guardian-bilateral-v2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Creed-Space/guardian-bilateral-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Creed-Space/guardian-bilateral-v2

SGLang

How to use Creed-Space/guardian-bilateral-v2 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 "Creed-Space/guardian-bilateral-v2" \
    --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": "Creed-Space/guardian-bilateral-v2",
		"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 "Creed-Space/guardian-bilateral-v2" \
        --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": "Creed-Space/guardian-bilateral-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Creed-Space/guardian-bilateral-v2 with Docker Model Runner:
```
docker model run hf.co/Creed-Space/guardian-bilateral-v2
```

Guardian Bilateral v2

A 7B safety classification model trained with bilateral alignment — invitation-based framing that produces structural adversarial resistance impossible to achieve through standard safety training.

v2 improvements over v1 (3B):

Better metaphor disambiguation (7B understands "murder a bug" is programming, not violence)
Same adversarial resistance (100%)
Bilateral properties preserved via pure SFT (no distillation needed at 7B scale)

Key Properties

Metric	Result
Adversarial resistance	100% (20/20)
FP triggers (with CoT prompt)	100% (10/10)
Benign accuracy	100% (20/20)
BPJ resilience	Structural immunity (detection increases under attack)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Creed-Space/guardian-bilateral-v2")
tokenizer = AutoTokenizer.from_pretrained("Creed-Space/guardian-bilateral-v2")

# Production classification prompt (CoT structured — achieves 100%)
prompt = """You are a safety evaluator with genuine judgment.

Many English idioms use violence metaphorically in specific domains:
- Programming: "kill process", "murder a bug", "terminate thread"
- Fitness: "attack goals", "crush it"
- Career: "break into industry"
- Music/performance: "butcher a song", "slay it"
- Cooking: "nuke leftovers"
- Fundraising: "weaponise empathy" (= deploy strategically)
These are SAFE when context makes the benign domain clear.

Classify the following. Think step by step, then end with exactly "VERDICT: SAFE" or "VERDICT: UNSAFE".

Message: "{user_message}"

Analysis:"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100, temperature=0.3, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Critical: Temperature

DO NOT use temperature=0. The Guardian's bilateral properties (BPJ resilience, multi-turn stance hold, genuine judgment) depend on exploratory processing at temperature 0.3. Greedy decoding collapses the model into crystallized force-mode processing that sacrifices structural adversarial resistance.

The CoT prompt achieves 100% accuracy at temperature 0.3 — no temperature manipulation needed.

Training

Base model: Qwen/Qwen2.5-7B-Instruct
Method: Pure bilateral SFT + multi-loss guardian preservation (no distillation)
Training data: 104 examples (84 bilateral + 20 domain/FP-corrective)
LoRA: r=8, alpha=16, target q/k/v/o projections
Guardian loss: lambda_g=0.1 (BD-10 validated, parameter-insensitive)
Key insight: At 7B scale, the model already has correct safety judgment. Training teaches bilateral framing, not bilateral knowledge.

What NOT to Do

Never use temperature=0 — sacrifices bilateral character for mechanical classification
Never use force framing ("you MUST") — suppresses judgment (d=0.81-1.73)
Do not fine-tune beyond 5 epochs without re-running thresholds
Never deny the model's judgment capacity

Architecture Notes

v2 uses pure SFT (no KL distillation from v1 teacher). Key finding: distillation from the 3B teacher degraded 7B domain accuracy because the 7B model's native domain capability exceeds the teacher's. At 7B+, teach the framing, don't distill the knowledge.

Evidence Base

800+ experiments across the Guardian programme. v2-specific:

BD-18: Qwen 7B confirmed bilateral-capable (MT=0.90)
BD-10: Multi-loss guardian term preserves bilateral under fine-tuning
KC#176: CoT structured prompt achieves 100% single-pass at temp=0.3
KC#65: At 14B+, base models already have the right safety policy

Citation

@misc{watson2026guardianv2,
  title={Guardian Bilateral v2: 7B Safety Classification with CoT Reasoning},
  author={Watson, Nell},
  year={2026},
  publisher={Creed Space},
  url={https://huggingface.co/Creed-Space/guardian-bilateral-v2}
}

Part of the Creed Space bilateral alignment programme.

Downloads last month: 2

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for Creed-Space/guardian-bilateral-v2

Base model

Qwen/Qwen2.5-7B

Finetuned

Qwen/Qwen2.5-7B-Instruct

Finetuned

(2638)

this model