sometimesanotion/Lamarck-14B-v0.3
Lamarck-14B-v0.3 by sometimesanotion is a 14.8 billion parameter language model developed through a multi-stage merge process, emphasizing both reasoning and prose quality. It integrates influences from models like arcee-ai/Virtuoso-Small and CultriX/SeQwence-14B-EvolMerge, with a context length of 32768 tokens. This model is designed for balanced performance in instruction following, reasoning, and generating high-quality prose.
Loading preview...
Lamarck-14B-v0.3 Overview
Lamarck-14B-v0.3 is a 14.8 billion parameter language model developed by sometimesanotion, distinguished by its sophisticated multi-stage merging strategy. This approach carefully balances influences from various base models to optimize for both strong reasoning capabilities and high-quality prose generation. The model leverages techniques like DELLA and SLERP merges to refine instruction following and achieve a balanced output.
Key Capabilities
- Balanced Performance: Engineered to provide a strong equilibrium between complex reasoning tasks and generating fluent, high-quality text.
- Advanced Merging Strategy: Utilizes a multi-stage merge process, including specialized branches for reasoning and prose, followed by DELLA and SLERP merges for refinement and balance.
- Influential Ancestors: Integrates top-performing models such as arcee-ai/Virtuoso-Small for prose quality and CultriX/SeQwence-14B-EvolMerge for reasoning.
- Instruction Following: Designed for smooth and accurate instruction adherence, benefiting from the careful balancing of its merged components.
Good For
- Applications requiring a strong blend of logical reasoning and creative or descriptive text generation.
- Tasks where both accurate instruction following and high prose quality are critical.
- Developers seeking a model built with advanced merging techniques to achieve specific performance characteristics.