Reproduce “Parameterized Synthetic Text Generation with SimpleStories” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lennart Finke, Chandan Sreedhara, Thomas Dooms, Mat Allen, Emerald Zhang, Juan Diego Rodriguez, Noa Nabeshima, Thomas Marshall and 1 more
Researchers at ETH Zürich and collaborators introduce SimpleStories, a parameterized synthetic text generation framework and dataset that improves upon TinyStories, yielding more diverse and labeled data for interpretability studies. Models trained on SimpleStories with a custom tokenizer achieve better performance on tasks like coherence and grammar with fewer parameters than prior benchmarks.
Key insights
- Parameterized prompting significantly enhances the diversity and label richness of synthetic datasets, making them more effective for training and interpretability tasks.
- Domain-optimized tokenizers are crucial for achieving high performance and parameter efficiency in small language models, often contributing more than model size alone.
- Mechanistic interpretability is greatly facilitated by clean, diverse, and labeled synthetic data, allowing for clearer probing of how models internally represent high-level linguistic and conceptual features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.