Hugging Face Chronos Gold 12B-1.0 Overview
Chronos Gold 12B-1.0, developed by elinas, is a unique 12 billion parameter model built upon the mistralai/Mistral-Nemo-Base-2407 architecture. It has undergone significant modifications and fine-tuning to enhance coherence and prompt adherence, making it comparable to larger models in specific applications.
Key Capabilities & Features
- General Chatbot Functionality: Designed to perform well in general conversational AI tasks.
- Roleplay & Storywriting: Optimized for creative text generation, including roleplay scenarios and story creation.
- Extended Sequence Generation: Observed to write sequences up to 2250 tokens in a single output.
- Context Length: Trained at a sequence length of 16384 tokens, maintaining the apparent 128k context length of Mistral-Nemo, though performance degrades beyond 16k.
- Character Card Support: Supports a majority of "character card" formats used in applications like SillyTavern.
- Enhanced Coherence: Re-creates the uniqueness of the original Chronos with significantly improved coherence and adherence to instructions.
Performance & Usage Notes
While the model retains the 128k context length from Mistral-Nemo, it is recommended to keep sequence lengths at a maximum of 16384 tokens to avoid performance degradation, as indicated by the RULER Test. The model uses the ChatML instruct template and is sensitive to high temperatures, with recommended sampling settings including a temperature of 0.7 (0.9 max) and a presence penalty of 1.0.
Open LLM Leaderboard Evaluation
Chronos Gold 12B-1.0 achieved an average score of 21.40 on the Open LLM Leaderboard, with notable scores in IFEval (31.66) and BBH (35.91). Detailed results are available here.