Matching domain experts by training from scratch on domain knowledge

TL;DR

Training small, domain-specific language models from scratch or fine-tuning them on specialized literature can achieve expert-level performance in neuroscience predictions, challenging the notion that larger models are necessary.

Contribution

Demonstrates that small LLMs trained specifically on neuroscience data can match expert performance, highlighting the importance of domain-specific training over model size.

Findings

Small models achieved expert-level neuroscience prediction performance.

Domain-specific training from scratch is effective for small LLMs.

Finetuning pretrained models on domain data also yields high performance.

Abstract

Recently, large language models (LLMs) have outperformed human experts in predicting the results of neuroscience experiments (Luo et al., 2024). What is the basis for this performance? One possibility is that statistical patterns in that specific scientific literature, as opposed to emergent reasoning abilities arising from broader training, underlie LLMs' performance. To evaluate this possibility, we trained (next word prediction) a relatively small 124M-parameter GPT-2 model on 1.3 billion tokens of domain-specific knowledge. Despite being orders of magnitude smaller than larger LLMs trained on trillions of tokens, small models achieved expert-level performance in predicting neuroscience results. Small models trained on the neuroscience literature succeeded when they were trained from scratch using a tokenizer specifically trained on neuroscience text or when the neuroscience literature was used to finetune a pretrained GPT-2. Our results indicate that expert-level performance may be attained by even small LLMs through domain-specific, auto-regressive training approaches.