File size: 1,631 Bytes
8e85a3e
 
 
82cab21
 
 
8e85a3e
82cab21
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
34
35
36
37
38
39
40
---
language:
- en
license: apache-2.0
pipeline_tag: text-generation
library_name: transformers
---

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

<p align="center">
    <a href="https://arxiv.org/abs/2508.19532">πŸ“„ Paper</a> β€’
    <a href="https://github.com/SenseLLM/StructureCoder">🏠 Repo</a> β€’
    <a href="https://huggingface.co/SenseLLM/StructureCoder-7B">πŸ€– 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 | πŸ€— [HF Link](https://huggingface.co/SenseLLM/StructureCoder-1.5B) | 1.5B |
| StructureCoder-3B   | πŸ€— [HF Link](https://huggingface.co/SenseLLM/StructureCoder-3B)   | 3B   |
| StructureCoder-7B   | πŸ€— [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)