qwen2.5-coder-3b-agent-v1

This repository contains a LoRA adapter, not a full standalone model.

It is the second-stage adapter in the project and was created by continuing fine-tuning from:

M-Alkassem/qwen2.5-coder-3b-unsloth-lora

The goal of this stage was to make the model more useful in a constrained tool-using workflow, especially for multi-step coding and debugging behavior.

What This Model Is

This adapter is the agent-oriented continued fine-tune in the project.

Training goal:

improve multi-step software-engineering behavior
improve inspect → reason → edit → test style behavior
make the model more useful inside a lightweight coding-agent loop

This adapter should be loaded on top of the Qwen2.5-Coder 3B base model.

Important Context

This adapter was not trained from scratch.

The training path was:

base model: unsloth/Qwen2.5-Coder-3B-Instruct-bnb-4bit
coding-focused adapter: M-Alkassem/qwen2.5-coder-3b-unsloth-lora
agent-oriented continued fine-tune: this repository

That means this adapter represents the latest learned state after both fine-tuning stages.

Dataset

This adapter was trained on a sampled subset of:

ernie-research/MEnvData-SWE-Trajectory

Project training setup:

sampled rows: 700
formatting strategy: tail-capped trajectory formatting to fit the token budget
max sequence length: 1024
training steps: 150

Training Summary

This model was trained with supervised fine-tuning (SFT) using LoRA and 4-bit quantization.

Key setup:

continued from the coding adapter
batch size per device: 1
gradient accumulation: 16
learning rate: 5e-5
optimizer: adamw_8bit
hardware: Google Colab Tesla T4

Observed result:

final training loss: about 1.2940

Intended Use

Use this adapter when you want:

a model that is better suited for a constrained coding-agent workflow
more agent-style behavior in inspect/edit/test tasks
a reasoning core for a lightweight tool-using coding agent

This adapter is most meaningful when paired with:

a controller loop
file tools
Python execution tools
iterative feedback from tool outputs

Limitations

This adapter is not a standalone merged model.

It also did not perform best in the plain direct-answer benchmark used in the project. In that evaluation, the original base model remained strongest overall.

So this adapter should not be presented as universally better at plain coding Q&A. Its value is more visible in tool-using and multi-step agent-style workflows.

How To Load

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

BASE_MODEL = "Qwen/Qwen2.5-Coder-3B-Instruct"
ADAPTER_MODEL = "M-Alkassem/qwen2.5-coder-3b-agent-v1"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
)

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, use_fast=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    torch_dtype=torch.float16,
    device_map="auto",
)

model = PeftModel.from_pretrained(base_model, ADAPTER_MODEL)
model.eval()

Example Prompt prompt = "A stack implementation fails a unit test when pop() is called on an empty stack. Explain how you would debug this step by step and propose a fix."

Project Context This adapter is part of a larger project with:

a coding-focused fine-tune an agent-oriented continued fine-tune a direct-answer benchmark comparing base vs coding adapter vs agent adapter a constrained agent_v2 prototype with file and Python tools In the documented agent_v2 run, the model was able to:

run failing tests detect a bug rewrite code rerun tests stop after success This is the main reason this adapter should be evaluated in both:

direct-answer mode tool-using agent mode References

Coding adapter: https://huggingface.co/M-Alkassem/qwen2.5-coder-3b-unsloth-lora
Base Qwen2.5-Coder model: https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct
Unsloth quantized base: https://huggingface.co/unsloth/Qwen2.5-Coder-3B-Instruct-bnb-4bit
Dataset card: https://huggingface.co/datasets/ernie-research/MEnvData-SWE-Trajectory

Citation

If you use this adapter, please cite the upstream Qwen2.5-Coder work and the dataset used for the agent-oriented continued fine-tune.

@article{hui2024qwen2p5coder,
  title={Qwen2.5-Coder Technical Report},
  author={Hui, Binyuan and Yang, Jian and Cui, Zeyu and Yang, Jing and Liu, Dayiheng and Zhang, Liqun and Liu, Tianyang and Zhang, Jiawei and Yu, Bo and Lu, Kaican and others},
  journal={arXiv preprint arXiv:2409.12186},
  year={2024}
}