---
license: llama2
---
<!-- markdownlint-disable first-line-h1 -->
<!-- markdownlint-disable html -->

<div align="center">
<h1>
  SlimPLM
</h1>
</div>

<p align="center">
📝 <a href="https://arxiv.org/abs/2402.12052" target="_blank">Paper</a> • 🤗 <a href="https://huggingface.co./zstanjj/SlimPLM-Query-Rewriting/" target="_blank">Hugging Face</a> • 🧩 <a href="https://github.com/plageon/SlimPLM" target="_blank">Github</a>
</p>

<div align="center">
</div>

## ✨ Latest News

- [1/25/2024]: Search Necessity Judgment Model released in [Hugging Face](https://huggingface.co./zstanjj/SlimPLM-Search-Necessity-Judgment/).
- [2/20/2024]: Query Rewriting Model released in [Hugging Face](https://huggingface.co./zstanjj/SlimPLM-Query-Rewriting/).

## 🎬 Get Started

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# construct prompt
question = "Who voices Darth Vader in Star Wars Episodes III-VI, IX Rogue One, and Rebels?"
heuristic_answer = "The voice of Darth Vader in Star Wars is provided by British actor James Earl Jones. He first voiced the character in the 1977 film \"Star Wars: Episode IV - A New Hope\", and his performance has been used in all subsequent Star Wars films, including the prequels and sequels."
prompt = (f"<s>[INST] <<SYS>>\nYou are a helpful assistant. Your task is to parse user input into"
          f" structured formats according to the coarse answer. Current datatime is 2023-12-20 9:47:28"
          f" <</SYS>>\n Course answer: (({heuristic_answer}))\nQuestion: (({question})) [/INST]")
params_query_rewrite = {"repetition_penalty": 1.05, "temperature": 0.01, "top_k": 1, "top_p": 0.85,
                        "max_new_tokens": 512, "do_sample": False, "seed": 2023}

# deploy model
model = AutoModelForCausalLM.from_pretrained("zstanjj/SlimPLM-Search-Necessity-Judgment").eval()
if torch.cuda.is_available():
    model.cuda()
tokenizer = AutoTokenizer.from_pretrained("zstanjj/SlimPLM-Search-Necessity-Judgment")

# run inference 
input_ids = tokenizer.encode(question, return_tensors="pt")
len_input_ids = len(input_ids[0])
if torch.cuda.is_available():
    input_ids = input_ids.cuda()
outputs = model.generate(input_ids)
res = tokenizer.decode(outputs[0][len_input_ids:], skip_special_tokens=True)
print(res)
```

## ✏️ Citation

```
@inproceedings{Tan2024SmallMB,
  title={Small Models, Big Insights: Leveraging Slim Proxy Models To Decide When and What to Retrieve for LLMs},
  author={Jiejun Tan and Zhicheng Dou and Yutao Zhu and Peidong Guo and Kun Fang and Jinhui Wen},
  year={2024},
  url={https://api.semanticscholar.org/CorpusID:267750726}
}
```