sometimesanotion/Lamarck-14B-v0.7

Lamarck-14B-v0.7: A Generalist Merge for Reasoning and Prose

Lamarck-14B-v0.7, developed by sometimesanotion, is a 14.8 billion parameter language model designed as a well-rounded generalist. This iteration has notably surpassed the 41.0 average maximum benchmark for 14B parameter models, demonstrating performance gains over previous releases.

Key Capabilities & Development

• Multi-step Reasoning: Engineered with a strong emphasis on complex problem-solving.
• Prose Generation: Excels in producing high-quality, nuanced text, benefiting from re-emphasis of models like Krystalan/DRT-o1-14B and underwoods/medius-erebus-magnum-14b.
• Multi-language Ability: Designed to handle multiple languages effectively.
• Advanced Merging Toolchain: Developed using a custom toolchain that automates LoRA and layer-targeting merges, incorporating techniques like SLERP+TIES for finalization.
• Ancestral Influence: Its performance is built upon a lineage of careful merges, integrating finetuning work from models such as arcee-ai/Virtuoso-Small, sthenno-com/miscii-14b-1225, and VAGOsolutions/SauerkrautLM-v2-14b-DPO for enhanced Big-Bench Hard (BBH) performance, alongside synergistic influences from DeepSeek-R1-Distill-Qwen-14B and others.

Good For

• Applications requiring robust multi-step reasoning.
• Tasks involving creative writing, content generation, or translation where high-quality prose is essential.
• General-purpose language understanding and generation where a balanced performance across various domains is desired.