Guilherme34/Firefly-V2

Firefly-V2 Overview

Firefly-V2 is a 3.2 billion parameter language model developed by Guilherme34. It is a merged model, combining the strengths of two distinct base models: Guilherme34/Firefly and SicariusSicariiStuff/Impish_LLAMA_3B. This merge was performed using the LazyMergekit tool and specifically employed the task arithmetic method.

Key Characteristics

• Model Architecture: A merge of two existing models, Guilherme34/Firefly and SicariusSicariiStuff/Impish_LLAMA_3B.
• Merge Method: Utilizes task_arithmetic for combining model weights, with Guilherme34/Firefly serving as the base model.
• Efficiency: Configured to use bfloat16 data type for potentially faster inference and reduced memory footprint.
• Context Length: Supports a substantial context window of 32768 tokens, allowing for processing longer inputs and generating more coherent extended outputs.

Usage and Application

Firefly-V2 is suitable for various text generation tasks. Its merged nature suggests an attempt to balance or enhance capabilities present in its constituent models. Developers can easily integrate Firefly-V2 into their projects using the Hugging Face transformers library, as demonstrated by the provided Python usage example for text generation.