AI BriefingStarCoder2-15B-Instruct Released
·2024.04.29 09:00
Key point
A code generation model that trained itself using StarCoder2-15B, without human intervention or paid LLMs, has been released.
Details
The HuggingFace BigCode team has released StarCoder2-15B-Instruct-v0.1. This model is the first fully transparent code LLM to apply a Self-alignment approach, in which the model generates its own training data without human annotations or data distillation from paid LLMs such as GPT-4.
Key Results and Features
- It achieved a HumanEval score of 72.6, surpassing CodeLlama-70B-Instruct (72.0).
- In LiveCodeBench tests, it outperformed models that used GPT-4 data. This suggests that a model can learn more effectively within its own data distribution than from a teacher model's data.
Data Generation Pipeline
- Extract high-quality, diverse seed functions from The Stack v1.
- Generate realistic instructions that incorporate coding concepts.
- Fine-tune the model using high-quality responses generated through execution-guided self-validation.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.