---
license: cc-by-nc-4.0
base_model:
- Qwen/Qwen2-1.5B-Instruct
tags:
- data processing
- slm
---

<div style="width: auto; margin-left: auto; margin-right: auto; background-color:black">
<img src="https://assets-global.website-files.com/6423879a8f63c1bb18d74bfa/648818d56d04c3bdf36d71ab_Refuel_rev8-01_ts-p-1600.png" alt="Refuel.ai" style="width: 100%; min-width: 400px; display: block; margin: auto;">
</div>


## Model Details

RefuelLLM-2-mini, aka Qwen-2-Refueled, is a Qwen-2-1.5B base model instruction tuned on a corpus of 2750+ datasets, spanning tasks such as classification, reading comprehension, structured attribute extraction and entity resolution. We're excited to open-source the model for the community to build on top of. 

More details about [RefuelLLM-2-mini](https://www.refuel.ai/blog-posts/refuel-llm-2-mini), and the [RefuelLLM-2 family of models](https://www.refuel.ai/blog-posts/announcing-refuel-llm-2). 

**Model developers** - Refuel AI

**Input** - Text only.

**Output** - Text only.

**Architecture** - Qwen-2-Refueled is built on top of a Qwen-2-1.5B base model. 

**Release Date** - Jan 8, 2025.

**License** - [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/deed.en)

## How to use

This repository contains weights for Qwen-2-Refueled that are compatible for use with HuggingFace. See the snippet below for usage with Transformers:

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

>>> model_id = "refuelai/Qwen-2-Refueled"
>>> tokenizer = AutoTokenizer.from_pretrained(model_id)
>>> model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

>>> messages = [{"role": "user", "content": "Is this comment toxic or non-toxic: RefuelLLM is the new way to label text data!"}]

>>> inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to("cuda")

>>> outputs = model.generate(inputs, max_new_tokens=20)
>>> print(tokenizer.decode(outputs[0]))
```

## Benchmarks

In this section, we report the output quality results on our benchmark of labeling tasks. For details on the methodology see [here](https://www.refuel.ai/blog-posts/refuel-llm-2-mini).


|      Model          |  Size | Overall   | Classification | Reading Comprehension | Structure Extraction | Entity Matching |
|---------------------|-------|-----------|----------------|-----------------------|-----------------------|-----------------|
| RefuelLLM-2-mini    | 1.5B  | **75.02%**| **72.18%**     | **78.18%**            | 75.18%                | 80.75%          |
| Qwen-2-3B           | 3B    | 67.62%    | 70.91%         | 71.53%                | **75.72%**            | 80.75%          |
| Phi-3.5-mini-instruct | 3.8B | 65.63%   | 70.57%         | 71.89%                | 65.34%                | **83.53%**      |
| Gemma-2-2B          | 2B    | 64.52%    | 67.99%         | 67.94%                | 76.01%                | 39.50%          |
| Llama-3-3B          | 3B    | 55.80%    | 55.81%         | 65.12%                | 61.50%                | 55.01%          |
| Qwen-2-1.5B         | 1.5B  | 51.22%    | 47.36%         | 67.15%                | 56.17%                | 45.25%          |
| Llama-3-1B          | 1B    | 39.92%    | 44.58%         | 29.67%                | 39.50%                | 62.94%          |



## Limitations

The Qwen-2-Refueled does not have any moderation mechanisms. We're looking forward to engaging with the community 
on ways to make the model finely respect guardrails, allowing for deployment in environments requiring moderated outputs.