SmolLM2-135M-Math

SmolLM2 Math

SmolLM2 but fine-tuned on math data!

I just made a fine-tune of SmolLM2 135M on the GSM8K dataset and it does improve math sometimes.

Evaluation Results

How to Use

This code is by Gemini 3 Flash:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "MihaiPopa-1/SmolLM2-135M-Math" # Replace with your repo path
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = "Question: If John has 5 apples and eats 2, then buys 4 more, how many does he have?\nAnswer:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Comparision

First Question

We both give the same question: "I have a number, 96. If the result is 24 and we divide it, what's the second part of the equation?"

SmolLM2 Math gave:

Question: I have a number, 96. If the result is 24 and we divide it, what's the second part of the equation?
Answer: The second part of the equation is 96 / 24 = 4.
#### 4
The answer is: 4

It correctly gave a answer.

SmolLM2 Base gave:

Question: I have a number, 96. If the result is 24 and we divide it, what's the second part of the equation?
Answer:

It produced NOTHING!

Second Question

Also, Gemini 2.5 Flash thought another question: "If 5 apples cost $2.50, how much does 1 apple cost?"

So we put this question in both models.

SmolLM2 Math gave:

Question: If 5 apples cost $2.50, how much does 1 apple cost?
Answer: 1 apple = 100 apples
2 apples = 100 * 2 = 200 apples
1 apple = 100 * 2 = 200 apples
#### 200
The answer is: 200

Wrong answer.

Second time:

Question: If 5 apples cost $2.50, how much does 1 apple cost? Think carefully.
Solution: 1/5 * $2.50 = $0.50
Therefore, 1 apple costs $0.50.
#### 0.50
The answer is: 0.50

Correct answer!

SmolLM2 Base gave:

Question: If 5 apples cost $2.50, how much does 1 apple cost?
Answer: 1 apple = $0.10

Reasoning: The answer is $0.10.

Wrong answer too.