NotHereNorThere/CoralLM-1b-raw
CoralLM-1b-raw is a 1 billion parameter language model developed by NotHereNorThere, created by merging three Llama-3.2-1B fine-tunes. This model combines capabilities in reasoning, mathematical/coding tasks, and creative writing, aiming for a generalist profile. It is designed to be a versatile 1B parameter model, integrating diverse strengths from its component models.
Loading preview...
CoralLM-1b-raw: A Merged 1B Parameter Generalist
CoralLM-1b-raw is a 1 billion parameter model developed by NotHereNorThere, created through a capability merge of three distinct Llama-3.2-1B fine-tunes. The primary goal of this merge was to combine specialized strengths in reasoning, math/coding, and creative writing into a single, more general-purpose 1B model.
Key Capabilities & Characteristics
- Merged Architecture: Utilizes the TIES merging method with
meta-llama/Llama-3.2-1B-Instructas its base, blending contributions from models focused on reasoning, math/coding, and general/creative tasks. - Diverse Skillset: Aims to integrate capabilities such as chain-of-thought (CoT) reasoning, mathematical problem-solving, coding logic, and creative writing.
- Raw Merge Output: This version represents the direct output of the merge process, without further training or cleanup.
- Context Length: Supports a context length of 32768 tokens.
Assessment & Considerations
While coherent, the model's responses can be scattered. It demonstrates strength in simple algebra and creative tasks. However, it may inappropriately reach for tools and over-explain. Due to its 1B parameter size, state-tracking problems are noted. The model inherits very weak alignment from a component, making it effectively uncensored and prone to attempting unsafe or harmful requests with little resistance.