How Magicoder takes LLM code generation to a new level
LLMs are very good at generating code. However, the highest-performing LLMs are the super-large, closed-source models like GPT-3.5 and GPT-4.
OSS-Instruct, a new technique developed by researchers at the University of Illinois at Urbana-Champaign and Tsinghua University, enables small language models to boost their coding capabilities to new levels and even approach the best-performing models.
OSS-Instruct generates fine-tuning examples in a way that is different from popular methods such as Self-Instruct and Evol-Instruct. It uses very small snippets of code to inspire a large model like GPT-4 to generate problems and solutions.
The researchers used OSS-Instruct to create Magicoder, a fine-tuned version of the 7-billion-parameter Code Llama that outperforms all other LLMs of similar size and other LLMs of much larger size. Interestingly, combining OSS-Instruct with Evol-Instruct takes Magicoder near the level of GPT-3.5 Turbo in generating code.
Read all about OSS-Instruct and Magicoder on TechTalks.
For more on AI research:
Optimize your ChatGPT prompts with DeepMind’s OPRO technique
Self-assembling neural networks can open new directions for AI research
LLMs can’t self-correct in reasoning tasks, DeepMind study finds
Recommendations:
My go-to platform for working with ChatGPT, GPT-4, and Claude is ForeFront.ai, which has a super-flexible pricing plan and plenty of good features for writing and coding. I use ForeFront for all kinds of tasks, including writing, coding, and testing new prompting techniques. The pricing is very convenient and the platform is user-friendly.