AuriAetherwiing/MN-12B-Starsong-v1
AuriAetherwiing/MN-12B-Starsong-v1 is a 12 billion parameter language model created by AuriAetherwiing, merged using the TIES method from nothingiisreal/MN-12B-Celeste-V1.9 and Sao10K/MN-12B-Lyra-v1. This model offers a 32768 token context length and appears to be more stable with Mistral formatting despite its base models being trained with ChatML. It is noted for being better suited for SFW content.
Loading preview...
AuriAetherwiing/MN-12B-Starsong-v1 Overview
MN-12B-Starsong-v1 is a 12 billion parameter language model developed by AuriAetherwiing, created through a merge of existing pre-trained models. It leverages the TIES merge method to combine nothingiisreal/MN-12B-Celeste-V1.9 as its base with Sao10K/MN-12B-Lyra-v1.
Key Characteristics
- Parameter Count: 12 billion parameters.
- Context Length: Supports a substantial 32768 token context window.
- Merge Method: Utilizes the TIES merge method for model integration.
- Formatting Stability: Noted to exhibit greater stability when used with Mistral formatting, despite its constituent models being trained with ChatML.
Intended Use Cases
- SFW Content Generation: The model is specifically highlighted as being more suitable for generating Safe-for-Work (SFW) content.
- Experimental Merges: Represents an exploration into model merging techniques, building upon experimental models like Sao10K's Lyra series.
Community-contributed GGUF and EXL2 quantizations are available for broader accessibility and deployment.