Hugging Face Sao10K/L3.1-70B-Euryale-v2.2 is a 70 billion parameter Llama 3.1-based model fine-tuned over two epochs for conversational instruction, creative writing, and roleplay. It features an extended context length of 32768 tokens. The model is specifically optimized for multi-turn coherency and creative text generation, incorporating datasets from human and Claude sources. It is designed for applications requiring nuanced roleplaying and creative narrative development.
Sao10K/L3.1-70B-Euryale-v2.2 Overview
This model is a 70 billion parameter variant based on the Llama 3.1 architecture, developed by Sao10K. It underwent a two-stage fine-tuning process, dedicating the first stage to multi-turn conversational instruction and the second to creative writing and roleplay tasks. The training utilized a diverse dataset combining human-generated content with data from Claude 3.5 Sonnet and Claude 3 Opus.
Key Enhancements & Capabilities
- Multi-turn Coherency: Significantly improved through the inclusion of dedicated multi-turn conversation-based instruct datasets.
- Creative Writing & Roleplay: Enhanced with 55% more roleplaying examples, drawing from Gryphe's Sonnet3.5-Charcard-Roleplay sets, and 40% more creative writing examples.
- Instruction Adherence: Includes datasets specifically targeting system prompt adherence and reasoning/spatial awareness.
- Refined Data Quality: Single-turn instruct data was replaced with higher-quality prompts and answers, and extensive filtering was performed to minimize errors.
Prompting Recommendations
Users are advised to use the Llama 3.1 Instruct Formatting, with the Euryale 2.1 Preset. Recommended inference parameters include a temperature of 1.2 and a min_p of 0.2 to achieve optimal creative and conversational outputs.