HuggingFaceTB/finemath
Viewer
•
Updated
•
48.3M
•
35.3k
•
241
FineMath datasets and ablation models
Note FineMath datasets
Note Llama 3B trained on a mix of FineMath and FineWeb-Edu: better at math and similar to Llama in reasoning, knowledge and common sense
Note FineMath text classifier to score the mathematical reasoning and educational content
Note Ablations on FineMath subsets (continual pre-training of base Llama 3.2 3B on 60B tokens)
Note Ablations on FineMath plus3 and plus4 (continual pre-training of base Llama 3.2 3B on 60B tokens)
Note Ablations on public math datasets and FW-Edu as a baseline (continual pre-training of base Llama 3.2 3B on 60B tokens)
Note Longer ablation for 160B on a mix of 40% fineweb-edu 60% FineMath and Infiwebmath 3plus / 4plus