Hugging Face's logo

SenseLLM
/
StructureCoder-3B

Text Generation
Transformers
Safetensors
English
qwen2
conversational
text-generation-inference
Model card Files
xet
Community
1
StructureCoder-3B
File size: 1,531 Bytes 
5b9cf69
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
 

## Alignment with Fill-In-the-Middle for Enhancing Code Generation

<p align="center">
    <a href="https://arxiv.org/abs/2508.19532">📄 Paper</a> •
    <a href="https://github.com/SenseLLM/StructureCoder">🏠 Repo</a> •
    <a href="https://huggingface.co/SenseLLM/StructureCoder-7B">🤖 Models</a>
</p>

## Introduction
Structure splits code snippets into smaller, granular blocks, creatingmore diverse DPO pairs from the same testcases. Additionally, we introduce the Abstract Syntax Tree (AST) splitting and curriculum training method to enhance the DPO training. Please refer to our paper for more details!

![](figures/method.png)

<hr>

## Models

| Model               | Checkpoint                                                        | Size |
|:--------------------|:------------------------------------------------------------------|:-----|
| StructureCoder-1.5B | 🤗 [HF Link](https://huggingface.co/SenseLLM/StructureCoder-1.5B) | 1.5B |
| StructureCoder-3B   | 🤗 [HF Link](https://huggingface.co/SenseLLM/StructureCoder-3B)   | 3B   |
| StructureCoder-7B   | 🤗 [HF Link](https://huggingface.co/SenseLLM/StructureCoder-7B)   | 7B   |

## Acknowledgments

We thank the following amazing projects that truly inspired us:

- [Qwen-Coder](https://github.com/QwenLM/Qwen3-Coder)
- [APPS](https://github.com/hendrycks/apps)
- [EvalPlus](https://github.com/evalplus/evalplus)
- [LiveCodeBench](https://github.com/LiveCodeBench/LiveCodeBench)
- [BigCodeBench](https://github.com/bigcode-project/bigcodebench)