TechFlow news — Recently, the full-stack data network Chainbase announced the open-sourcing of Theia-Llama-3.1-8B, its large language model specifically designed for the crypto domain, on HuggingFace. The model outperforms mainstream models in both perplexity and BERT scores, demonstrating superior understanding of the crypto world compared to most leading open-source large models.
The Chainbase team has uniquely constructed the first specialized Web3 dataset, compiling various materials from the top 2000 projects listed on CoinMarketCap. This dataset has undergone both manual and algorithmic filtering to ensure accuracy, diversity, and domain expertise in the training data. Leveraging this dataset, the team efficiently fine-tuned the model using LoRA technology and accelerated training with tools such as DeepSpeed. Additionally, the model has been quantized into Q8 GGUF format, significantly reducing memory usage and improving inference speed.
Notably, Theia-Llama-3.1-8B represents Chainbase’s initial effort in developing large-scale models for the cryptocurrency space. The model has already been successfully deployed in Chainbase's interactive demo application, TheiaChat, which currently boasts over 300,000 daily active users.
