BlueNipples/SnowLotus-v2-10.7B

SnowLotus-v2-10.7B: Enhanced Prose and Intelligence

SnowLotus-v2-10.7B is a 10.7 billion parameter model developed by BlueNipples, created through a strategic slerp merge. The primary goal of this model is to combine strong prose generation with high intelligence, specifically targeting uncensored roleplay scenarios.

Key Components and Capabilities

• Core Ingredients: The model is built upon Sao10K's Frostwind finetune, known for its intelligence, and NyxKrage's Frostmaid experiment, which excels in prose and creative long-range context handling.
• Merging Strategy: It incorporates NyxKrage's solardoc into Frostwind, then gradient SLERP'd the result into Frostmaid, with specific filter values applied to self_attn and mlp layers.
• Context Length: While the base Solar model is designed for 4k context, components of this merge have shown functionality up to 8k tokens.
• Improved Prose: Compared to previous versions, SnowLotus-v2-10.7B offers better prose, less "GPT-ish" language, and maintains coherency.

Performance and Use Cases

• Ayumi Index Scores: In the Ayumi ERPv4 Chat Log Index, SnowLotus scores 94.10 in Flesch readability, indicating more complex sentence structures. It also scores 70.94 on IQ4, outperforming DaringLotus and ranking highly within its weight class.
• Descriptive Ability: Both SnowLotus and DaringLotus demonstrate strong descriptive abilities, with SnowLotus ranking highly for adjective and adverb usage.
• Roleplay Optimization: The model is specifically designed for roleplaying, aiming to provide a smarter version of Nyx's Frostmaid with enhanced prose.
• Potential Repetition: Benchmarking suggests potential for repetition, though users are advised to use modern samplers like MinP or DynaTemp to mitigate this.