FORGE-Qwen-Final

🚀 A lightweight coding model adapter fine-tuned for FORGE-v4, an adversarial self-improvement environment for robust code generation.

---

🔥 Model Overview

This repository contains a LoRA / PEFT adapter trained on top of:

Base Model: Qwen/Qwen2.5-Coder-1.5B-Instruct

The adapter was trained in Google Colab GPU using:

• Unsloth
• Hugging Face TRL
• Supervised Fine-Tuning (SFT)
• FORGE robustness tasks

---

🛡️ Purpose

Most coding models solve normal tasks but fail on edge cases.

This adapter improves reliability for:

• Negative numbers
• Duplicate values
• Empty arrays
• Safe sorting tasks
• Clean Python function formatting
• Instruction following

---

🧠 About FORGE-v4

FORGE is an adversarial benchmark + training system where:

• Defender model writes code
• Breaker agent generates hidden edge cases
• Reward system scores robustness
• Training improves failure cases

Theme Alignment:

• Self-Improvement
• Multi-Agent AI
• Verifiable RL Environments

---

⚙️ Training Details

Hardware: Google Colab Tesla T4 GPU Method: 4-bit QLoRA fine-tuning Adapter Type: LoRA Training Stack: Unsloth + TRL

Example metrics:

• Training completed successfully
• Exported real adapter weights
• Public deployment ready

---

📦 Files Included

• adapter_model.safetensors
• adapter_config.json
• tokenizer files

---

🚀 How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base = "Qwen/Qwen2.5-Coder-1.5B-Instruct"
adapter = "sanjay7676/forge-qwen-final"

tokenizer = AutoTokenizer.from_pretrained(base)

model = AutoModelForCausalLM.from_pretrained(base)

model = PeftModel.from_pretrained(model, adapter)

---

💻 Example Prompt

Return only Python code.

Write:

def solve(arr):

Sort integers preserving duplicates and negatives.

---

🌐 Used In

FORGE-v4 live environment / Hugging Face Space demo.

---

📈 Future Work

• GRPO reinforcement tuning
• More adversarial coding tasks
• Runtime efficiency rewards
• Multi-language coding benchmarks

---

👤 Author

Sanjay7676

Built for FORGE-v4 hackathon submission.