DarkSapling-7B-v2.0 Overview
DarkSapling-7B-v2.0 is a 7 billion parameter language model developed by TeeZee, created by merging four distinct Mistral-7B based models: dolphin-2.6-mistral-7b-dpo-laser, Mistral-7B-Holodeck-1, Mistral-7B-Erebus-v3, and samantha-mistral-7b. This version utilizes the DARE TIES merging method, which aims to better preserve the characteristics of each constituent model compared to previous iterations.
Key Capabilities
- Enhanced Roleplay: Specifically designed for one-on-one ERP, demonstrating improved empathy and the ability to handle both SFW and NSFW content seamlessly.
- Character Adherence: Excels at sticking to provided character cards, ensuring consistent persona generation.
- Storytelling: Offers satisfactory storytelling capabilities, influenced by the Holodeck model in its merge.
- Instruction Following: Shows good proficiency in following instructions.
- Context Switching: Capable of smoothly transitioning between different conversational contexts.
Performance Metrics
Evaluated on the Open LLM Leaderboard, DarkSapling-7B-v2.0 achieved an average score of 64.98.
- AI2 Reasoning Challenge (25-Shot): 64.16
- HellaSwag (10-Shot): 85.10
- MMLU (5-Shot): 64.37
- TruthfulQA (0-shot): 52.21
- Winogrande (5-shot): 78.61
- GSM8k (5-shot): 45.41
Noteworthy Characteristics
This model is described as more romantic and empathetic than its v1.0 counterpart, and generally smarter. It can produce dark scenarios due to the influence of Erebus and is noted for its ability to generate both SFW and NSFW content without issues. Users should be aware that the model can produce NSFW content.