You can deploy this model and start earning money today!

qqqq43w34dfgd/
SmolLM2-1.7B-Instruct-QLoRA

Name: SmolLM2-1.7B-Instruct-QLoRA
Availability: InStock

2502.02737

2412.08347

Views

No views yet

SmolLM2 Instruct — Train Your Own AI for FREE

🚀 The most efficient open-source AI you can train for free — even on CPU!

Two Training Options

⭐ Option A: CPU Training (No GPU needed!)

Uses SmolLM2-135M — a tiny 135M parameter model that trains on any CPU.

Component	Details
Base Model	SmolLM2-135M (135M params)
Dataset	Capybara (16k multi-turn conversations)
Method	LoRA (trains only ~2.5M params — 50x faster!)
Hardware	Google Colab CPU (free, default!)
Training Time	~4-6 hours
Script	`train_colab_cpu.py`

Option B: GPU Training (better quality)

Uses SmolLM2-1.7B — the best sub-2B model, needs a free T4 GPU.

Component	Details
Base Model	SmolLM2-1.7B (1.7B params)
Dataset	smol-smoltalk (500k+ samples)
Method	QLoRA (4-bit quantization + LoRA r=32)
Hardware	Google Colab T4 GPU or Kaggle T4
Training Time	~6-8 hours
Script	`train_colab.py`

Papers: SmolLM2, SmolTulu

🏃 How to Train (CPU — Recommended for Free Users)

Go to Google Colab
Create a new notebook (default CPU runtime is fine!)
Copy the contents of train_colab_cpu.py into a cell
Replace YOUR_HF_TOKEN with your token from huggingface.co/settings/tokens (needs Write access)
Replace YOUR_HF_USERNAME with your Hugging Face username
Click Run ▶ and wait ~4-6 hours
Your trained model auto-pushes to your HF account! 🎉

Quick Start (after training)

python

from transformers import pipeline

pipe = pipeline("text-generation", model="YOUR_USERNAME/SmolLM2-135M-Instruct")
messages = [{"role": "user", "content": "Explain quantum computing in simple terms"}]
response = pipe(messages, max_new_tokens=256)
print(response[0]["generated_text"][-1]["content"])

What your model will be able to do

💬 Follow instructions and have conversations
📝 Help with writing, summarization, and editing
❓ Answer questions on many topics
🧮 Basic math and reasoning

Why these models?

SmolLM2 models outperform similar-sized competitors on most benchmarks
LoRA makes training 50x faster by only updating ~2% of parameters
Capybara dataset is diverse, high-quality, and small enough for CPU training
Research-backed hyperparameters from published papers (SmolTulu)

License

Apache 2.0 (same as base models)

qqqq43w34dfgd/SmolLM2-1.7B-Instruct-QLoRA