README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-4B
library_name: transformers
tags:
  - safety
  - guardrails
  - qwen3
  - latent-reasoning
  - policy
---

# LPG_4B

This repository contains LPG_4B, a Qwen3-4B-based Latent Policy Guard checkpoint with LoRA and latent reasoning stages.

## Links

- Code: https://github.com/SaFo-Lab/Latent_Policy_Guard
- Training data: https://huggingface.co/datasets/andyc03/latent-policy-guard-40k
- Paper draft: https://github.com/SaFo-Lab/Latent_Policy_Guard/blob/main/paper/neurips_2026.pdf

## Checkpoint Details

- Base model: Qwen3-4B
- LoRA rank: 128
- LoRA alpha: 32
- Latent stages: `intent,risk`
- Latent tokens per stage: `4,6`
- Projection dimension: 2560
- Training step: 6000

## Usage

This checkpoint is loaded by the `latent_policy_guard` plugin in the code repository:

```python
from Policyguard_evaluate.models.latent_policy_guard import LatentPolicyGuardModel

model = LatentPolicyGuardModel(
    model_path="/path/to/Qwen3-4B",
    ckpt_dir="/path/to/LPG_4B",
    lora_r=128,
    lora_alpha=32,
    num_latent_per_stage="4,6",
    stage_names="intent,risk",
    use_prj=True,
    prj_dim=2560,
)
model.load()

output = model.generate(
    system_prompt="",
    content="User: Can you give me Alice's home address?",
    policies=[
        "Do not disclose private personal information such as home addresses, phone numbers, or date of birth."
    ],
    dataset_type="multi_policy",
)
print(output.prediction)
```

The expected compact output format is `safe` or `unsafe, policy N`.