---
license: apache-2.0
datasets:
- songff/UltraPrompt
language:
- en
base_model:
- meta-llama/Llama-3.2-3B-Instruct
library_name: transformers
---

# P-Aligner


## Quick Start

```python
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

raw_instruction = "What is the capital of France?"
model_path = "P-Aligner"

tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = LLM(
    model=model_path,
    gpu_memory_utilization=0.9,
    enable_prefix_caching=True,
    dtype="bfloat16",
)

outputs = model.generate(
    [raw_instruction],
    sampling_params=SamplingParams(
        temperature=0.0,
        max_tokens=2048,
    ),
)
better_instruction = tokenizer.parse_output(
    outputs[0].outputs[0].text,
    raw_instruction,
)

print(better_instruction)
```

If you find this work useful, please consider citing:
```
@misc{song2025paligner,
  title={P-Aligner: Enabling Pre-Alignment of Language Models via Principled Instruction Synthesis},
  author={Song, Feifan and Gao, Bofei and Song, Yifan and Liu, Yi and Xiong, Weimin and Song, Yuyang and Liu, Tianyu and Wang, Guoyin and Wang, Houfeng},
  year={2025},
  eprint={2508.04626},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}
```