Golden-Curry-12B: A Specialized Roleplaying LLM
Golden-Curry-12B is a 12 billion parameter language model from GreenerPastures, engineered specifically for advanced roleplaying scenarios. Built upon the Mistral NeMo base, it features a 32768 token context window, enabling sustained and complex character interactions. The model's tokenizer was modified for ChatML compatibility prior to pretraining, ensuring seamless integration with standard chat interfaces.
Key Capabilities & Training
- Immersive Roleplaying: Excels at maintaining consistent character personas, dynamic narrative generation, and crafting emotionally resonant dialogue.
- ChatML Compatibility: Fully supports ChatML format due to tokenizer modifications and reinforced training.
- Specialized Training: Underwent a multi-stage training process including custom pretraining on narrative fiction, instruct tuning, and a final roleplaying tune. A Kahneman-Tversky optimization was applied for alignment.
- Diverse Data Foundation: Trained on a wide array of instruction and roleplaying datasets, including Mielikki/Erebus-87k, PocketDoc/Dans-MemoryCore-CoreCurriculum-Small, and Epiculous/SynthRP-Gens-v1.1-Filtered-n-Cleaned, among others.
Ideal Use Cases
- Interactive Fiction: Generating rich, branching narratives where the model plays a character.
- Character Simulation: Creating virtual characters that can engage in realistic and persona-consistent conversations.
- Chat-based Roleplay: Providing an engaging and dynamic partner for user-driven roleplaying sessions.
Quantized GGUF versions are available for optimized deployment.