Overview
TINY_SCHILLER is a German literary text corpus designed for small language models, bridging the gap in prototyping, fine-tuning, education, and research. It offers a single-file solution similar to Karpathy's tiny_shakespeare.
Features of the Corpus
- Size: 2.07 megabytes, containing eleven public-domain Schiller dramas.
- Source: Sourced from DraCor's GerDraCor export (CC0), processed by deterministic parser engineering.
- Encoding: Utilizes character-level, GPT-2 byte-pair encoding, and cl100k_base tokenization splits.
- Characteristics: Includes instruction-formatted dialogue-completion splits and 89 per-character persona splits.
Code Example
The corpus can be loaded with a single call from HuggingFace, as shown in the code below:
// Load German literary text for small language model
model.load("tiny_schiller");This design allows small language models to access German literary text with minimal code, significantly facilitating research and applications.
Blogger's Review: The release of TINY_SCHILLER not only simplifies access to German literary texts but also provides great convenience for applications of small language models. Its one-line loading feature showcases the efficiency and ease of modern NLP, making it worthy of attention.