SmolLM3-3B-MathReason

A math-focused fine-tuned version of SmolLM3-3B, optimized for step-by-step mathematical reasoning and problem solving.

Highlights

📚 Math-First: Trained on ~7K high-quality math and reasoning samples

🧠 Chain-of-Thought: Supports /think mode for detailed reasoning

⚡ Lightweight: 3B parameters, runs on consumer GPUs

Training Details

Parameter	Value
Base Model	HuggingFaceTB/SmolLM3-3B
Method	LoRA (r=16, alpha=32)
Training Data	~7K samples
- OpenThoughts3_1.2M_think	5,000 (math reasoning)
- s1k_1.1_think	~1,000 (high-quality math)
- smoltalk_everyday_convs	1,000 (everyday reasoning)
Epochs	2
Learning Rate	2e-4 (cosine)
Effective Batch Size	16

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("real-jiakai/SmolLM3-3B-MathReason")
tokenizer = AutoTokenizer.from_pretrained("real-jiakai/SmolLM3-3B-MathReason")

messages = [
    {"role": "system", "content": "/think"},  # Enable reasoning mode
    {"role": "user", "content": "A store sells apples for $2 each. If John buys 5 apples and pays with a $20 bill, how much change does he get?"}
]

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

Intended Use

• GSM8K style math problems
• Step-by-step problem solving
• Educational math tutoring
• Arithmetic and algebra reasoning

Limitations

• English only
• May struggle with very complex multi-step problems
• Not designed for factual knowledge retrieval

Training Infrastructure

• GPU: NVIDIA A100
• Training Time: ~2 hours
• Framework: TRL + PEFT