Lamarck-14B-v0.6: A Generalist Merge for Reasoning and Prose

Lamarck-14B-v0.6 is a 14.8 billion parameter language model developed by sometimesanotion, designed as a generalist merge with a strong focus on multi-step reasoning, prose generation, and multi-language ability. It has achieved the #1 average score on the Open LLM Leaderboard for text-generation assistant models under 32 billion parameters, validating its complex merge methodology.

Key Capabilities & Features

• Advanced Merging Techniques: Utilizes a custom toolchain for merges, including extracted LoRA adapters, separate branches for breadcrumbs and DELLA merges, highly targeted weight/density gradients across layers, and finalization through SLERP merges. This approach selectively combines the strongest aspects of its ancestral models.
• Reasoning Focus: Incorporates contributions from models like Krystalan/DRT-o1-14B, which is based on research into Deep Reasoning Translation via Long Chain-of-Thought.
• Generalist Performance: Aims to provide a balanced model that excels across various tasks, particularly in complex reasoning and high-quality text generation.

When to Use This Model

• Complex Reasoning Tasks: Ideal for applications requiring multi-step logical deduction or problem-solving.
• High-Quality Prose Generation: Suitable for creative writing, content generation, or any task demanding nuanced and well-structured text.
• Multi-language Applications: Benefits from its multi-language capabilities for diverse linguistic needs.