The Wave Field LLM (v4) model has demonstrated effective scalability up to 825 million parameters, approaching the billion threshold.

Training Details

The model training took 13.2 hours on a dataset of 1.33 billion tokens, reaching a final perplexity of 72.2 and an accuracy of 27.1%. These results indicate that the model is stable, converges correctly, and effectively handles large volumes of tokens.

Implications

The success of Wave Field LLM validates the field-based approach as a promising interaction mechanism for language models. This opens up new possibilities for the development of alternative architectures to traditional transformers, potentially more efficient in terms of computation and memory.