The final agent adapter was trained by continuing from the coding adapter, so this merged model represents the latest learned state after both fine-tuning stages.

Training Background

Stage 1: Coding Fine-Tune

Dataset:

bigcode/self-oss-instruct-sc2-exec-filter-50k

Setup:

sampled rows before filtering: 4000
rows used after filtering: 3993
max sequence length: 1024
training steps: 250

Result:

final training loss: about 0.6130

Stage 2: Agent-Oriented Continued Fine-Tune

Dataset:

ernie-research/MEnvData-SWE-Trajectory

Setup:

sampled rows: 700
max sequence length: 1024
training steps: 150

Result:

final training loss: about 1.2940

Evaluation Notes

In the direct-answer benchmark, the original base model remained the strongest plain answer-only model overall.

The main value of this final merged model is different:

it is the final standalone artifact of the project
it is more aligned to constrained tool-using workflows
it performed best when used as the reasoning core of a lightweight coding agent

The benchmark summary image above shows the plain prompting comparison:

Base model overall mean: 3.97
Coding adapter overall mean: 2.97
Agent adapter overall mean: 1.77

The agent workflow image shows the documented agent_v2 result where the model:

ran failing tests
identified a bug
rewrote code
reran tests
stopped after success

How To Load

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "M-Alkassem/qwen2.5-coder-3b-final-merged"

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

model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto",
)

model.eval()

Related Repositories

coding adapter: M-Alkassem/qwen2.5-coder-3b-unsloth-lora
agent adapter: M-Alkassem/qwen2.5-coder-3b-agent-v1

References

Base model: https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct
Coding adapter: https://huggingface.co/M-Alkassem/qwen2.5-coder-3b-unsloth-lora
Agent adapter: https://huggingface.co/M-Alkassem/qwen2.5-coder-3b-agent-v1
Coding dataset: https://huggingface.co/datasets/bigcode/self-oss-instruct-sc2-exec-filter-50k
Agent dataset: https://huggingface.co/datasets/ernie-research/MEnvData-SWE-Trajectory

Citation

If you use this model, please cite:

bibtex

@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}
}