Raspberry-Pie/Hot-Coco

🥤 Hot Coco: Grab a Hot Coco and Start Coding

Never train an AI model again. Hot Coco is your personal, on-demand expert for PyTorch, Hugging Face Transformers, and Machine Learning.

Built on Mistral-7B and fine-tuned with QLoRA, Hot Coco skips the fluff and gets straight to the code. Whether you are debugging a CUDA error or architecting a new Transformer, just grab a Hot Coco and start coding.

🚀 Why Hot Coco?

• 🎯 Specialized Expertise: Trained exclusively on high-quality PyTorch and ML data.
• ⚡ Lightweight & Fast: Uses 4-bit quantization (QLoRA) to run smoothly on consumer hardware.
• 🧠 Long Context: Handles up to 32k tokens, perfect for pasting entire scripts for debugging.
• 🔒 Private & Secure: Keep your proprietary code safe by running it locally.

🛠️ Quick Start

1. Install Dependencies

pip install transformers peft accelerate bitsandbytes torch

2. Load Hot Coco

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
from huggingface_hub import login

# Authenticate for private repo access
login(token="YOUR_HF_TOKEN")

base_model_id = "mistralai/Mistral-7B-Instruct-v0.2"
adapter_id = "Raspberry-Pie/Hot-Coco"

# Configure 4-bit loading for efficiency
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16
)

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    device_map="auto"
)

# Merge Hot Coco's expertise
model = PeftModel.from_pretrained(base_model, adapter_id)

3. Start Coding

prompt = "<s>[INST] Write a PyTorch training loop with gradient clipping. [/INST]"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📝 Training Details

• Base: Mistral-7B-Instruct-v0.2
• Method: QLoRA (4-bit NF4)
• Data: 1,800 curated examples from CodeAlpaca + Custom Identity Injection
• Hardware: Trained on Kaggle T4 x2 GPUs

---

Built with ❤️ by Raspberry-Pie