QUICKSTART.md

Wraith Coder 7B

Signal-dense code generation model fine-tuned from Qwen2.5-Coder-7B-Instruct.

Quick Start

Installation

pip install transformers torch

Basic Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "vanta-research/wraith-coder-7b",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("vanta-research/wraith-coder-7b")

messages = [
    {"role": "user", "content": "Implement binary search with complexity analysis."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Ollama Deployment

# Convert to GGUF (Q4_K_M recommended)
ollama create wraith-coder:7b -f Modelfile

# Run inference
ollama run wraith-coder:7b "Implement a LRU cache with O(1) operations"

Key Features

• 62.6% more concise than base Qwen2.5-Coder-7B while maintaining correctness
• 60% complexity analysis coverage across diverse coding challenges
• Multiple solution approaches with trade-off discussions
• Systems programming knowledge integrated throughout
• Production-ready for senior engineering applications

Performance Highlights

Metric	Base Qwen	Wraith Coder	Improvement
Avg Response Length	2,900 chars	1,084 chars	62.6% shorter
Complexity Analysis	40%	60%	+50% coverage
Multiple Approaches	35%	65%	+86% frequency
Trade-off Discussion	45%	75%	+67% depth

Documentation

Full documentation available in README.md

License

Apache 2.0

Citation

@misc{wraith-coder-7b,
  author = {Vanta Research},
  title = {Wraith Coder 7B: Signal-Dense Code Generation through Iterative Fine-Tuning},
  year = {2025},
  publisher = {Hugging Face}
}