Surge AI’s Post

The most powerful LLMs in the world are trained on Surge AI’s RLHF. Some of these models include code LLMs which are going to be huge for progress in AI! Let’s look at some of the recent developments in code LLMs: Foundational Code LLMs - Research indicates that foundational code LLMs could be vital for enabling powerful code understanding and code generation capabilities necessary in more advanced applications of LLMs that range from personalized code assistants to AI-powered debugging tools. What makes a good code LLM? It’s really about the data quality. As with other types of general-purpose LLMs, high-quality data is key to training code LLMs. Let's look at some recent developments in code LLMs to find out more. Code Llama - There is a lot we can learn from the recently released code LLM called Code Llama. Code Llama leverages code-specific datasets but there is also a variant called Code Llama - Instruct that leverages proprietary instruction tuning and self-instruct datasets used in Llama 2 to inherit instruction following, helpfulness, and safety properties. Code Generation Results - Other code LLMs like AlphaCode and StarCode are trained using code only while Code Llama leverages the foundational Llama 2 model. These models are evaluated on description-to-code generation benchmarks. The code-heavy datasets are key to enhanced results in code generation. In fact, even the Code Llama 7B compares or outperforms Llama 70B on Python coding benchmarks and multilingual evaluation. Red Teaming - Similar to general-purpose LLMs, code LLMs can also benefit from red teaming which involves identifying risks through adversarial prompting in the context of coding. For instance, it can help to prevent the LLM from generating malicious code. While this is beneficial for high-stakes applications, it can result in LLMs over-refusing which might lead to bad user experience in some domains. It’s clear to see how careful attention to building high-quality datasets can help in training state-of-the-art code LLMs, enabling their utility, and getting desired capabilities. If you need help with training your code LLMs, red-teaming, or collecting high-quality datasets, reach out to our team: https://lnkd.in/eGiZPbub