🔬 Code Autopsy

A QLoRA fine-tuned code review model based on Qwen2.5-Coder-7B-Instruct.
Paste buggy code → get a structured analysis: bug ID, root cause, and auto-fix.

🤗 Model  ·  🧩 LoRA Adapter  ·  🚀 Live Demo  ·  📊 W&B Run

---

🖥️ Interactive Review Interface

Code Autopsy Diagnostic Report: Automated bug identification (mutable default argument), root cause breakdown, 92.3% confidence score, and optimized code generation.

Real-Time Inference State: Live neural diagnostics runner analyzing Python source code with animated status spinner.

---

Architecture

flowchart LR
    A["🐛 Buggy Code\n(Python / JS)"] --> B["FastAPI\n/review"]
    B --> C["Qwen2.5-Coder-7B\n+ LoRA Adapter\n4-bit NF4"]
    C --> D["Structured Output"]
    D --> E1["🐛 Bug Identified"]
    D --> E2["🔍 Root Cause"]
    D --> E3["✅ Fixed Code"]
    D --> E4["📊 Confidence"]

    F["Gradio UI\n(HF Spaces)"] -->|"POST /review\nvia API_URL"| B
    G["Ollama\n(local GGUF)"] -->|alternative| C

Loading

---

Results

Quantitative: sacreBLEU (n=20)

Model	BLEU Score	Δ vs Base
Base (Qwen2.5-Coder-7B)	10.63	—
Code Autopsy (fine-tuned)	70.03	+59.40 (+558%)

Fine-tuned in 1 epoch on an RTX 5060 (8 GB VRAM) using 4-bit NF4 QLoRA.

Training Convergence

Metric	Value
Final train loss	0.407
Final eval loss	0.297
Eval token accuracy	92.3%
Trainable params	5.05M / 4.36B (0.12%)

---

Before / After Examples

Python — ZeroDivisionError

	Code
Buggy	def avg(nums): return sum(nums) / len(nums)
Fixed	def avg(nums): return sum(nums) / len(nums) if nums else 0.0

Bug Identified: ZeroDivisionError when nums is an empty list.
Root Cause: No guard clause for the empty input case. len([]) returns 0, causing division by zero.

---

Python — Mutable Default Argument

	Code
Buggy	def append(val, lst=[]): lst.append(val); return lst
Fixed	def append(val, lst=None): if lst is None: lst = []; lst.append(val); return lst

Bug Identified: Mutable default argument lst=[] is shared across all calls.
Root Cause: Python evaluates default arguments once at function definition time, not per-call.

---

JavaScript — Missing await

	Code
Buggy	const data = response.json(); return data.name;
Fixed	const data = await response.json(); return data.name;

Bug Identified: response.json() returns a Promise, not the parsed data.
Root Cause: Response.json() is asynchronous — without await, data holds the Promise object.

---

JavaScript — var Closure in Loop

	Code
Buggy	for (var i = 0; i < 3; i++) { setTimeout(() => console.log(i), 100); }
Fixed	for (let i = 0; i < 3; i++) { setTimeout(() => console.log(i), 100); }

Bug Identified: Prints 3, 3, 3 instead of 0, 1, 2.
Root Cause: var is function-scoped; all closures share the same i. Replace with let (block-scoped).

---

Quick Start (Windows)

1. Double-click INSTALL.bat    ← sets up venv + installs everything
2. Fill in your keys in .env
3. Double-click Run_Project.bat ← starts API + UI, opens browser

---

Local Setup (Manual)

1. Clone & install

git clone https://github.com/Ares19v/Code-Autopsy.git
cd Code-Autopsy

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate

# Install PyTorch first (match your CUDA version)
pip install torch --index-url https://download.pytorch.org/whl/cu124

pip install -r requirements.txt

2. Configure environment

cp .env.example .env
# Fill in: HF_TOKEN, WANDB_API_KEY, GEMINI_API_KEY

3. Prepare dataset

# Dry-run first (no downloads, validates formatting logic)
python data/prepare_dataset.py --dry-run

# Full run (~20k examples, downloads ~5GB)
python data/prepare_dataset.py --max-samples 20000

4. Train

# Debug: 1 step, no W&B (validates VRAM fits)
python training/train.py --debug

# Full training run (1 epoch, logs to W&B)
python training/train.py

5. Evaluate

# sacreBLEU: base vs fine-tuned
python eval/eval_codebleu.py --adapter ./adapter --max-samples 20

# LLM-as-judge (100 examples, calls Gemini API)
python eval/eval_llm_judge.py --adapter ./adapter

6. Serve (FastAPI)

uvicorn serve.api:app --host 0.0.0.0 --port 8000

# Test it:
curl -X POST http://localhost:8000/review \
  -H "Content-Type: application/json" \
  -d '{"code": "def avg(x): return sum(x)/len(x)", "language": "python"}'

7. Launch Gradio demo

python demo/app.py
# Opens at http://localhost:7860

---

Docker

# Build and run everything (API + Demo)
docker compose up --build

# API only
docker build -f serve/Dockerfile -t code-autopsy-serve .
docker run -p 8000:8000 \
  -v $(pwd)/adapter:/app/adapter:ro \
  --env-file .env \
  --gpus all \
  code-autopsy-serve

---

Ollama (local GGUF)

# 1. Merge adapter + convert to GGUF
python publish.py --skip-merge=false   # creates ./merged_model/
bash convert_to_gguf.sh               # requires llama.cpp

# 2. Create & run
ollama create code-autopsy -f Modelfile
ollama run code-autopsy

---

Publish to HuggingFace

# Push both merged model + raw adapter
python publish.py

# Adapter only (faster, no merge step)
python publish.py --skip-merge

---

Project Structure

code-autopsy/
├── data/
│   ├── raw/                    # gitignored
│   ├── processed/              # gitignored
│   └── prepare_dataset.py
├── training/
│   ├── train.py                # SFTTrainer + QLoRA main script
│   ├── config.yaml             # All hyperparameters
│   └── utils.py                # Shared helpers
├── eval/
│   ├── eval_codebleu.py        # Quantitative BLEU eval
│   ├── eval_llm_judge.py       # Gemini LLM-as-judge
│   ├── plot_results.py         # Results chart generator
│   └── results/                # Saved evaluation outputs
├── serve/
│   ├── api.py                  # FastAPI /review endpoint
│   └── Dockerfile
├── demo/
│   ├── app.py                  # Gradio UI
│   └── Dockerfile
├── tests/
│   └── test_utils.py           # CI unit tests (no GPU required)
├── .github/
│   └── workflows/ci.yml        # GitHub Actions CI
├── publish.py                  # Merge + push to HF Hub
├── Modelfile                   # Ollama config
├── convert_to_gguf.sh          # GGUF conversion
├── docker-compose.yml          # Full stack orchestration
├── requirements.txt
├── .env.example
├── INSTALL.bat                 # Windows one-click setup
├── Run_Project.bat             # Windows one-click launcher
└── README.md

---

Hardware

Trained on an HP Omen with RTX 5060 using 4-bit NF4 quantization via BitsAndBytes.

Setting	Value
GPU	NVIDIA RTX 5060 (8 GB VRAM)
Quantization	NF4 4-bit (BitsAndBytes)
Precision	bfloat16 compute
Effective batch	16 (4 × 4 grad accum)
Max seq length	512 tokens
Optimizer	adamw_8bit
Epochs	1

---

Devansh Tyagi @ 2026

🤗 Model on Hugging Face

The trained model is available on Hugging Face: devanshty/Code-Autopsy

Download

from huggingface_hub import hf_hub_download
model_path = hf_hub_download(repo_id='devanshty/Code-Autopsy', filename='adapter_model.safetensors')

---

© 2026 Devansh Tyagi (Ares19v). All Rights Reserved.