File size: 2,148 Bytes
ba84daf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
41
42
43
44
45
46
47
48
49
50
51
52
53
54
---

license: other
license_name: qwen-research
license_link: LICENSE
base_model: huihui-ai/Qwen2.5-3B-Instruct-abliterated-SFT
base_model_relation: finetune
language:
  - en
library_name: transformers
pipeline_tag: text-generation
tags:
  - python
  - code-generation
  - lora
---


# Python Coder 3B Research

**Built with Qwen.** This 3.09B-parameter model completes short Python functions from
a function signature and a docstring. It is prepared for noncommercial research and
evaluation under the [Qwen Research License](LICENSE).

## Training

This model started from [huihui-ai/Qwen2.5-3B-Instruct-abliterated-SFT](https://huggingface.co/huihui-ai/Qwen2.5-3B-Instruct-abliterated-SFT)
at revision e21343fbcabbb6225d1cb9122061404f2d21f1ad, derived from
[Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct).
The Python fine-tune used 120 LoRA steps on 165 synthetic training rows. The LoRA
updates were merged into the BF16 weights. The four SafeTensors shards contain
the modified model weights. See [NOTICE](NOTICE) for attribution and file changes.

## Evaluation

A prompt passed when its generated function passed every test input. These were
small, synthetic Python suites scored using a restricted subset of the language.

| Suite | Source checkpoint | This model |
| --- | ---: | ---: |
| Familiar operation patterns | 124/144 | 125/144 |
| Original combination challenge | 24/24 | 24/24 |
| Fresh task families | 27/40 | 24/40 |

The 60% fresh-task score applies to short functions in this evaluation. The
benchmark did not measure repository-level bug fixing, other programming languages,
or refusal behavior. Detailed results are in [evaluation.json](evaluation.json).

## Inference

Load this repository with Transformers using AutoTokenizer and
AutoModelForCausalLM. Use the included chat template. To match the benchmark
format, give the signature and docstring in the user message and place the same
signature and docstring at the start of the assistant completion. The model then
generates the function body. Review and test generated code before using it.