qwen2.5-coder-3b-unsloth-lora

qwen2.5-coder-3b-unsloth-lora

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

It was created by fine-tuning unsloth/Qwen2.5-Coder-3B-Instruct-bnb-4bit for coding-assistance behavior on Google Colab using T4 GPU.

What This Model Is

This adapter is the first-stage coding-focused fine-tune in the project.

Training goal:

• improve structured coding responses
• improve instruction-following for programming tasks
• improve simple bug-fixing behavior

This adapter should be loaded on top of the base model:

• base model: unsloth/Qwen2.5-Coder-3B-Instruct-bnb-4bit

Dataset

This adapter was trained on a sampled subset of:

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

Project training setup:

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

Training Summary

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

Key setup:

• LoRA rank: 16
• LoRA alpha: 16
• LoRA dropout: 0
• batch size per device: 1
• gradient accumulation: 16
• learning rate: 1e-4
• optimizer: adamw_8bit
• hardware: Google Colab Tesla T4

Observed result:

• final training loss: about 0.6130

Intended Use

Use this adapter when you want:

• a lightweight coding assistant
• better structured code answers
• simple debugging help
• a PEFT adapter that runs on top of the Qwen2.5-Coder 3B base model

Limitations

This adapter is not a standalone merged model.

It also was not the strongest model in the later direct-answer benchmark on every prompt. It improved some focused coding-task behavior, but it should be understood as a practical low-resource experiment rather than a universally superior model.

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-unsloth-lora"

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 = "Debug this Python code and explain the bug: def is_even(n): return n % 2 == 1"

Project Context This adapter is part of a larger two-stage project:

coding-focused adapter: this repository agent-oriented continued adapter: M-Alkassem/qwen2.5-coder-3b-agent-v1 The later agent adapter was trained by continuing from this coding adapter.

References

• Qwen2.5-Coder base 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/bigcode/self-oss-instruct-sc2-exec-filter-50k