Purpose
CodeGen2 is the official research model series released by Salesforce for the task of program synthesis. Program synthesis means letting the model automatically generate code based on a given context. The related research was published as a paper at ICLR 2023, titled "CodeGen2: Lessons for Training LLMs on Programming and Natural Languages," with authors including Erik Nijkamp, Hiroaki Hayashi, Caiming Xiong, Silvio Savarese, and Yingbo Zhou, among whom the first two authors are marked as having equal contribution.
Main capabilities
This series provides models at four parameter scales: 1B, 3B, 7B, and 16B. Model checkpoints have been released on the Hugging Face Hub, and the model cards explain how to use the models for causal sampling and infill sampling. Among these, program synthesis can be implemented through auto-regressive sampling.
Usage
According to official materials, users can obtain the model checkpoints through the Hugging Face Hub and refer to the corresponding model cards to learn the specific usage of causal sampling and infill sampling. Auto-regressive sampling can be used to perform program synthesis tasks. Apart from this, the materials do not provide more detailed installation commands, code examples, or operating steps; this content is not yet provided.
Applicable scenarios
CodeGen2 is mainly oriented toward research and development scenarios in code generation and program synthesis, and is suitable for related work that requires code generation experiments based on open-source models and research on methods for training large language models on programming languages and natural languages.
Limitations
The existing materials only state that this is an official research release and do not mention specific information such as charging models, pricing, performance results, or commercial licensing; this content is not yet provided.
