| # Linear Next Benchmark | |
| Linear Next is a comprehensive benchmark designed to fairly compare various efficient transformer architectures. This project evaluates different approaches including linear attention, sparse attention, and other model structures under identical training conditions and datasets. | |
| ## Overview | |
| The benchmark aims to provide an unbiased comparison of efficient transformer variants by ensuring all models are trained with the same datasets, hyperparameters, and evaluation metrics. This allows for a clear understanding of the relative strengths and weaknesses of each approach. | |
| ## Datasets | |
| The benchmark utilizes a diverse collection of high-quality datasets: | |
| ### General Text | |
| - **DCLM-pro**: A large-scale dataset containing diverse text from various domains, designed for general language modeling tasks. | |
| - **Cosmopedia-v2**: A curated corpus of high-quality web content covering a wide range of topics, with emphasis on educational and informative material. | |
| - **Fineweb-edu**: A filtered collection of educational web content, focusing on instructional and academic text from reliable sources. | |
| ### Code | |
| - **The Stack v2**: A comprehensive collection of source code spanning multiple programming languages, designed to train models on code understanding and generation tasks. | |
| ### Mathematics | |
| - **Finemath**: A specialized dataset containing mathematical content, including equations, proofs, and mathematical explanations across various difficulty levels. | |
| ### Reasoning | |
| - **Natural Reasoning**: A dataset focused on logical reasoning, problem-solving, and inference tasks, designed to improve models' reasoning capabilities. | |
| ## Methodology | |
| All models in the Linear Next benchmark are evaluated using identical: | |
| - Training datasets and data mixing ratios | |
| - Optimization parameters | |
| - Hardware configurations | |
| - Evaluation metrics | |
| This controlled environment ensures that performance differences can be attributed to the architectural differences rather than training conditions. | |
| ## Results | |
| Detailed benchmark results, including training curves, inference speed, memory usage, and performance metrics across different tasks, are available in the project repository. | |