Neph0s/CoSER-Llama-3.1-70B
Neph0s/CoSER-Llama-3.1-70B is a 70 billion parameter language model developed by Neph0s, fine-tuned from LLaMA-3.1. It is specifically designed for role-playing language agents (RPLAs), excelling at generating human-like responses for diverse personas, including fictional and original characters. Trained on the CoSER dataset of authentic multi-turn dialogues from novels, it demonstrates state-of-the-art performance in maintaining character consistency and adapting to complex role-playing scenarios.
Loading preview...
CoSER-Llama-3.1-70B: Advanced Role-Playing Language Agent
Neph0s/CoSER-Llama-3.1-70B is a 70 billion parameter model built upon the LLaMA-3.1 base, specifically engineered for role-playing language agents (RPLAs). It is part of the CoSER model family, which focuses on generating highly human-like and consistent character portrayals across various scenarios.
Key Capabilities and Differentiators
- Exceptional Role-Playing: CoSER models are trained on the unique CoSER dataset, which features authentic multi-turn, multi-character dialogues extracted from 771 renowned novels. This enables the model to capture nuanced personalities, maintain consistent character traits, and adapt to diverse role-playing contexts.
- Comprehensive Training Data: The CoSER dataset includes character profiles, dialogues, plot summaries, character experiences, and conversation backgrounds, along with internal thoughts and physical actions, leading to richer character portrayals.
- Given-Circumstance Acting (GCA) Methodology: The training approach optimizes the language modeling loss for each character's messages within a given conversation, characters, and setting, enhancing the model's ability to sequentially portray characters accurately.
- State-of-the-Art Performance: CoSER-70B outperforms existing open-source LLMs on multiple RPLA benchmarks, demonstrating superior performance in metrics like Anthropomorphism (53.33), BLEU (10.10), and ROUGE-L (14.78) in GCA Evaluation. It also achieves leading scores on existing RPLA benchmarks such as Life Choice (93.47) and CroSS MR (64.49), often comparable to closed-source models like GPT-4o.
Ideal Use Cases
This model is particularly well-suited for applications requiring:
- Interactive Storytelling and Narrative Generation: Creating dynamic and character-driven stories.
- Virtual Companions and Chatbots: Developing agents that can maintain consistent personas.
- Game Development: Populating games with NPCs that exhibit believable and engaging personalities.
- Creative Writing Assistance: Aiding writers in developing character dialogues and interactions.