This model experiment was inspired by the work published in Goat: Fine-tuned LLaMA Outperforms GPT-4 on Arithmetic Tasks, which found good success in fine tuning Llama models on math.
Fine tuning of philschmid/Llama-2-7b-hf was conducted with 2.8M math problems from the AtlasUnified/atlas-math-sets dataset.
Training was conducted on a trn1.32xlarge instance. The model here was compiled for 2 Neuron cores, which will run on AWS inf2.8xlarge and larger instances.
The table below highlights performance improvements observed on basic math questions.
Question | Correct Answer | Llama2-7B Answer | llama2-7B-AtlasMath-2.8M |
---|---|---|---|
What is the addition of 591 + 642 | 1,233 | $354$ | 1,233 |
Multiply these numbers: 3 * 12 | 36 | If you multiply the numbers in an array by one another, the resulting number is called the product of those numbers. For example, 3 * 12 = 36. | 36 |
Can you determine the result of 100 / 12 | 0.833 | 16.66666666666667 ### Solution First we can write 100 as 100 * 0.1 Next we can write 12 as 12 * 0.1 Next we can write | 0.8 |
Can you solve this equation: 9 * 7 | 63 | ### Instruction Can you solve this equation: 2 * 9 ### Answer 18 | 63 |
I need the sum of 3000 + 2635 | 5,635 | ### Output 2635 ### Error You have an error in your SQL syntax; check the manual that corresponds to your MySQL server version for the right syntax to use near '+ 2635' at line 1 | 5,635 |
Can you figure out the cube root of 99 | 4.626 | 11.85 | 99.85 |
What do you get when you add 8711 + 5691 | 14,402 | 6402 ### Source Advent of Code | 14,402 |
Multiply 123 * 123 | 15,129 | 55,609 ### Hint 1 What's the difference between this expression and this one? 123 * 123 = 14,999 | 15,069 |
Please compute the square root of 10 | 3.162 | 3.160 | 3303.88 |
Please find the sum of 1000 + 52 | 1,052 | 352 ### Solution python def solution(n): return sum(map(lambda x: x + 52, range(n))) |
6256 |
Subtract 3 from 28938 | 28,935 | -1260 ### Explanation This numeric expression can be rewritten as 28937 + (-1) - (-1) and evaluated as 28938 - 3 . ### Notes - The -3 does not |
28935 |
To use the model:
from optimum.neuron import pipeline
p = pipeline('text-generation', model="hartmani/llama2-7B-AtlasMath-2.8M", tokenizer="philschmid/Llama-2-7b-hf")
p("What is the addition of 591 + 642", max_new_tokens=64, do_sample=True, top_k=50)
Disclaimer: This model was generated for research purposes only and may produce inconsistent or inaccurate results. There are obviously far better ways to have computers perform basic math calculations. This model simply demonstrates the ease of teaching a Llama2-7B model basic math techniques.
- Downloads last month
- 12