LINAGORA and the OpenLLM-France consortium have released Luciole-23B-Instruct-1.1, a fine-tuned and aligned version of their open-source multilingual causal language model.
The model was trained on Jean Zay in three phases: supervised fine-tuning on instruction data with thinking traces, supervised fine-tuning without thinking traces, and preference alignment using Direct Preference Optimization (DPO). The training data covers math, science, coding, general chat, RAG, and translation. Smaller 8B and 1B variants are also available.
This release provides an Apache 2.0 licensed option for multilingual instruction following, supported by the France 2030 program.