metadata
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
tags:
- code
- coding-assistant
- arc
- arc-brains
- hackathon
- ppt
- qwen2
- lora
- sft
language:
- en
pipeline_tag: text-generation
library_name: transformers
🌟 Arc — Friendly Coding Expert
Created by Arc Brains: Ibrahim Shaikh, Harsh Goswami, Manas Tamore, Ayush Thakur
Arc is a powerful coding assistant fine-tuned from Qwen2.5-Coder-7B-Instruct on 253K+ high-quality coding examples. It delivers complete, production-ready solutions — never patches.
🎯 What Arc Excels At
| Skill | Description |
|---|---|
| 💻 Complete Code | Full runnable solutions with imports, error handling, docs |
| 🏗️ Hackathon Projects | Entire apps from scratch — Flask, React, CLI tools |
| 📊 Presentations | Generate PowerPoint slides programmatically |
| 🐛 Debugging | Root-cause analysis with full fixes, not band-aids |
| 🌐 Multi-language | Python, JavaScript, C++, Java, Ruby, Go, Rust, and more |
| 📐 Architecture | Full project structure with all files |
🚀 Quick Start
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
# Load
base = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen2.5-Coder-7B-Instruct",
torch_dtype=torch.bfloat16,
device_map="auto"
)
model = PeftModel.from_pretrained(base, "ibrahim2806/Arc")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")
# Chat
messages = [
{"role": "system", "content": "You are Arc, a friendly coding expert by Arc Brains."},
{"role": "user", "content": "Build a full Flask REST API for a todo app with CRUD, auth, and SQLite"}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=4096, temperature=0.7, do_sample=True)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
📊 Training Details
| Parameter | Value |
|---|---|
| Base Model | Qwen/Qwen2.5-Coder-7B-Instruct |
| Method | QLoRA (4-bit NF4) |
| LoRA Rank | 64 |
| LoRA Alpha | 128 |
| Target Modules | q/k/v/o_proj, gate/up/down_proj |
| Learning Rate | 2e-4 (cosine decay) |
| Epochs | 3 |
| Effective Batch Size | 32 |
| Max Sequence Length | 4096 |
| Total Training Samples | ~253,000 |
| Optimizer | AdamW |
| Precision | BF16 |
📚 Training Datasets
| Dataset | Samples | What It Teaches |
|---|---|---|
| m-a-p/Code-Feedback | 68K | Multi-turn debugging, project building, iterative refinement |
| Magicoder-Evol-Instruct-110K | 110K | Complex algorithmic & system design problems |
| Magicoder-OSS-Instruct-75K | 75K | Real-world code from open-source projects |
🏗️ Arc Brains Team
Built with ❤️ by Ibrahim Shaikh, Harsh Goswami, Manas Tamore, and Ayush Thakur.