SelfCodeAlign
paper Your tags
Your notes
Fully transparent, permissive self-alignment for code LLMs (NeurIPS 2024): the model generates its own instruction data (concepts → tasks → responses, validated by self-generated tests) with no human annotation and no distillation from larger proprietary models. Produced StarCoder2-Instruct, the first fully self-aligned open code LLM. UIUC-led (Yuxiang Wei, Lingming Zhang) within the BigCode consortium, with Snowflake and Hugging Face — joint BigCode attribution; StarCoder2 itself is ServiceNow/HF-led and not booked here.