Curating LLM Tuning Data from the FineWeb Dataset for High-fidelity Domain Adaptation

We created a post-training dataset from FineWeb dataset for high-fidelity domain adaptation of open weight LLM (Google Flan). Parameter efficient fine-tuning through prompt tuning resulted in remarkable improvement in perplexity scores as well as demonstration of ability of the tuned model to generalize based on information in the tuning dataset.

The work was selected for oral presentation at AGU24. Slide attached.