GOAT
About
The GOAT family of LLMs, developed by the GOAT.AI research lab, is dedicated to enhancing human-AI interactions and addressing common challenges faced by large language models, such as hallucinations and context comprehension issues. Their flagship model, GOAT-7B-Community, is a refined iteration of Meta's LLaMA 2 7B model, meticulously trained on 72,000 multi-turn dialogues sourced from the GoatChat app and OpenAssistant 45. This model showcases exceptional capabilities within the open-source 7B model category, excelling in benchmarks like MMLU and BigBench Hard 5. Despite these advancements, the relatively small model size of 7 billion parameters presents challenges such as the occurrence of hallucinations 5. The GOAT.AI team is actively working to mitigate these issues through experiments with larger models and improved data cleansing methods 5. Alongside GOAT-7B-Community, the lab has introduced other specialized models like the GOAT-70B-Storytelling, intended for narrative generation tasks 3, reflecting a focus on data quality and innovative training techniques to boost performance and safety 5.