AI21 Labs Challenges GPT-3 with New Jurassic-1 Language Models
AI21 Labs introduces Jurassic-1, a pair of massive auto-regressive language models designed to compete directly with OpenAI's GPT-3. The company releases the models alongside a new open-source evaluation suite for testing zero-shot and few-shot learning capabilities.
AI21 Labs officially releases Jurassic-1, a duo of auto-regressive language models that directly rival GPT-3. The suite includes J1-Jumbo, a massive 178B-parameter model, and J1-Large, a smaller 7B-parameter model designed for more agile tasks.
Researchers evaluate both models using perplexity metrics alongside zero-shot and few-shot learning benchmarks. These tests reveal how effectively the Jurassic-1 models perform against GPT-3 when generating text without extensive prior training on specific prompts.
To support broader industry testing, AI21 Labs publicly releases a specialized zero-shot and few-shot evaluation suite on GitHub. This open-source tool provides a shared resource that allows developers to accurately assess the performance of mega-scale language models moving forward.