ToyFT 500
A toy dataset for testing fine tuning with 500 samples. Derived from MATH-500 with an added messages
column for compatibility with fine tuning libraries. I've tested full-finetuning with gemma-3-270m-it and it can overfit successfully in about 10 epochs!
IMPORTANT! Please do not use this outside of simple testing. MATH-500 is still used for benchmarking today so fine tuning on this dataset will lead to contamination.
- Downloads last month
- 18