L3.1 70B Euryale v2.2

L3.1 70B Euryale v2.2 is Sao10K's creative roleplay model, fine-tuned on Meta's Llama 3.1 foundation with 70 billion parameters and a 131K context window suited to extended roleplay and narrative-heavy work. Training ran over two epochs, with the first focused on multi-turn conversational instruction and the second on creative writing and roleplay data, drawing on a mix of human and Claude-generated examples. Compared with v2.1, this release adds noticeably more roleplay and creative writing examples, along with datasets targeting reasoning and spatial awareness. The result is better system prompt adherence and steadier multi-turn coherence for immersive character interaction and world-building. Use it for conversational agents, creative applications, and interactive storytelling that benefit from a long context and consistent character dynamics. Sao10K suggests a temperature around 1.2 with min_p near 0.2 and the Llama 3 Instruct format.