---
base_model: unsloth/Qwen3-4B-Instruct-2507
library_name: peft
license: apache-2.0
tags:
  - qlora
  - lora
  - unsloth
  - trl
  - agentic
  - tool-use
  - code
  - reasoning
  - calibration
datasets:
  - nvidia/Open-SWE-Traces
  - open-thoughts/OpenThoughts3-1.2M
  - Agent-Ark/Toucan-1.5M
language:
  - en
pipeline_tag: text-generation
---

# Luck-Qwen3-4b-Code-FineTune

A QLoRA fine-tune of `unsloth/Qwen3-4B-Instruct-2507` specialized in agentic coding, multi-step reasoning, tool use, and calibrated uncertainty.

**TL;DR:** This model improves output formatting consistency, tool interaction, and knowledge boundary awareness ("I don't know" responses) without increasing the underlying capability ceiling of the base model.

<p align="center">
  <img src="training_loss.png" alt="Training Loss Curve" width="100%">
</p>
<p align="center"><em>Training loss convergence from ~1.7 to ~0.55 across 1,267 steps.</em></p>

## Model Summary

| Parameter / Metric | Value |
|---|---|
| **Base Model** | `unsloth/Qwen3-4B-Instruct-2507` (4.05B params, dense) |
| **Method** | QLoRA (4-bit NF4) |
| **LoRA Rank / Alpha** | r=16, alpha=16, dropout=0 |
| **Target Modules** | `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj` |
| **Trainable Parameters** | 33,030,144 / 4,055,498,240 (0.81%) |
| **Context Length** | 4,096 tokens |
| **Dataset Size** | 10,129 examples |
| **Training Steps / Epochs** | 1,267 steps / 1 epoch |
| **Effective Batch Size** | 8 (2 per-device × 4 gradient accumulation steps) |
| **Optimizer / Schedule** | `adamw_8bit`, learning rate 2e-4, cosine schedule |
| **Hardware** | Single T4 GPU (Google Colab / Kaggle) |
| **Final Loss** | ~0.55 (started at ~1.7) |
| **Frameworks** | Unsloth, TRL (`SFTTrainer`), PEFT, Transformers |

## In This Repo

This repository contains:
- **QLoRA LoRA Adapter weights**
- **Merged FP16 safetensors weights**

*Note: Quantized GGUF files are hosted separately at [`ahmetggg/Luck-Qwen3-4b-Code-FineTune-GGUF`](https://huggingface.co/ahmetggg/Luck-Qwen3-4b-Code-FineTune-GGUF).*

## Intended Use

- **Agentic Coding:** Repository-level navigation, issue resolution, and code generation.
- **Multi-Step Reasoning:** Step-by-step problem-solving in science and math domains.
- **Tool Use:** Structured multi-turn function and tool calling.
- **Calibrated Uncertainty:** Refusing to answer ("I don't know") when encountering actual knowledge gaps.

## Training Data

| Source | Focus Area | Details / Filtering |
|---|---|---|
| [`nvidia/Open-SWE-Traces`](https://huggingface.co/datasets/nvidia/Open-SWE-Traces) | Agentic coding | Filtered exclusively to `resolved == 1` trajectories |
| [`open-thoughts/OpenThoughts3-1.2M`](https://huggingface.co/datasets/open-thoughts/OpenThoughts3-1.2M) | Math & science reasoning | Code domain excluded (handled by Open-SWE-Traces) |
| [`Agent-Ark/Toucan-1.5M`](https://huggingface.co/datasets/Agent-Ark/Toucan-1.5M) | Tool calling | Real multi-turn tool interaction trajectories |
| **Custom R-Tuning Set** | Calibration | ~150–300 MMLU items queried on base model; correct outputs retained as normal QA, incorrect outputs relabeled as "I don't know" |

## Usage

### Option 1: Unsloth (LoRA Adapter)

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "ahmetggg/Luck-Qwen3-4b-Code-FineTune",
    max_seq_length = 4096,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

messages = [{"role": "user", "content": "Write a Python function to reverse a linked list."}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))