---
library_name: transformers
language:
- en
base_model:
- meta-llama/Llama-3.1-70B-Instruct
tags:
- evaluation
---

<div align="center">

# LMUnit: Fine-grained Evaluation with Natural Language Unit Tests
<img src="Contextual_AI_Brand_Mark_Dark.png" width="10%" alt="Contextual_AI"/>

</div>

<hr>

<div align="center">

[![Paper](https://img.shields.io/badge/Paper-LMUnit-blue)](https://arxiv.org/abs/2412.13091)
[![Blog Post](https://img.shields.io/badge/📝%20Blog-LMUnit-green)](https://contextual.ai/research/lmunit)
[![GitHub](https://img.shields.io/badge/GitHub-LMUnit-black?logo=github)](https://github.com/ContextualAI/LMUnit)
[![Hugging Face Collection](https://img.shields.io/badge/🤗%20Hugging%20Face-Model%20Collection-yellow)](https://huggingface.co/collections/ContextualAI/lmunit)

</div>

**LMUnit** is a state-of-the-art language model that is optimized for evaluating natural language unit tests. It takes three inputs: a prompt, a response, and a unit test. It then produces a continuous score between 1 and 5 where higher scores indicate that the response better satisfies the unit test criteria.

The LMUnit model achieves leading averaged performance across preference, direct scoring, and fine-grained unit test evaluation tasks, as measured by FLASK and BiGGen Bench, and performs on par with frontier models for coarse evaluation of long-form responses (per LFQA). The model also demonstrates exceptional alignment with human preferences, ranking in the top 5 of the RewardBench benchmark with 93.5% accuracy and in top #2 of RewardBench2 with 82.1% accuracy.

For more details, please check out the [blogpost](https://contextual.ai/research/lmunit) or the [paper](https://arxiv.org/abs/2412.13091).

## Model Details

LMUnit is highly performant and versatile because of key methodologies in its training approach:

- **Multi-Objective Training:** The model simultaneously learns from multiple evaluation signals, including pairwise comparisons between responses, direct quality ratings, and specialized criteria-based judgments.
- **Synthetic Data Generation:** We developed a sophisticated pipeline to generate training data that captures nuanced, fine-grained evaluation criteria and subtle quality distinctions between responses across a wide range of use cases and scenarios.
- **Importance Weighting:** We demonstrate that adjusting unit test weights to reflect the relative importance of different criteria achieves results that better align with human preferences.

### Model Description

- **Developed by:** Contextual AI
- **Language(s) (NLP):** English
- **Finetuned from model:** Llama-3.1-70B-Instruct

### Model Sources

- **Repository:** https://github.com/ContextualAI/LMUnit
- **Paper:** https://arxiv.org/abs/2412.13091

## 🚀 Model Quick Start

### Installation
```bash
pip install lmunit
```

### Basic Usage
```python
from lmunit import LMUnit
from vllm import SamplingParams

# Initialize LMUnit
model = LMUnit(
    model_path="ContextualAI/LMUnit-llama3.1-70b",
    tp_size=4
)

# Define evaluation
query = "What is the capital of France?"
response = "Paris"
unit_test = "Does the response correctly identify the capital city?"

# Generate score
sampling_params = SamplingParams(temperature=0.0, max_tokens=10, logprobs=20)
prompt = f"Query: {query}\n\nResponse: {response}\n\nUnit Test: {unit_test}"
output = model.generate(prompt, sampling_params)
print(output)
```

### Alternative: Using Transformers
```python
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load model
tokenizer = AutoTokenizer.from_pretrained("ContextualAI/LMUnit-llama3.1-70b")
model = AutoModelForCausalLM.from_pretrained("ContextualAI/LMUnit-llama3.1-70b")

# Prepare prompt
query = "What is the capital of France?"
response = "Paris"
unit_test = "Does the response correctly identify the capital city?"
content = f"Query: {query}\n\nResponse: {response}\n\nUnit Test: {unit_test}"

messages = [{"role": "user", "content": content}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

# Generate
outputs = model.generate(**inputs, max_new_tokens=40)
result = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])
print(result)
```

For more examples, see our [GitHub repository](https://github.com/ContextualAI/LMUnit).

### Evaluation - Results

| Model | Flask | BiGGen-Bench | Human-Internal | InfoBench | RB | LFQA | RB2 |
|:------|------:|-------------:|---------------:|----------:|----:|------:|----:|
| **LMUnit-LLaMA-3.1-70B** | 72.03 | 67.69 | 93.63 | 89.00 | 91.56 | 76.15 | 80.5 |
| **LMUnit-Qwen2.5-72B** | 73.85 | 69.56 | 94.44 | 88.67 | 91.13 | 73.85 | 82.1 |

## Citation

If you find our work helpful, feel free to cite our paper:
```bibtex
@inproceedings{saadfalcon2025lmunit,
      title={{LMUnit}: Fine-grained Evaluation with Natural Language Unit Tests}, 
      author={Jon Saad-Falcon and Rajan Vivek and William Berrios and Nandita Shankar Naik and Matija Franklin and Bertie Vidgen and Amanpreet Singh and Douwe Kiela and Shikib Mehri},
      booktitle={Findings of the Association for Computational Linguistics: EMNLP 2025},
      year={2025},
      url={https://arxiv.org/abs/2412.13091}
}
```