darkc0de/BuddyGlassKilledBonziBuddy

TEXT GENERATIONConcurrency Cost:2Model Size:24BQuant:FP8Ctx Length:32kPublished:Feb 26, 2025Architecture:Transformer Cold

darkc0de/BuddyGlassKilledBonziBuddy is a 24 billion parameter language model created by darkc0de, merged using the TIES method from a Mistral-Small-24B-Base-2501 foundation. This model integrates capabilities from TheDrummer/Cydonia-24B-v2, arcee-ai/Arcee-Blitz, and cognitivecomputations/Dolphin3.0-Mistral-24B. With a 32768 token context length, it is designed to combine the strengths of its constituent models for diverse applications.

Loading preview...

Overview

darkc0de/BuddyGlassKilledBonziBuddy is a 24 billion parameter language model developed by darkc0de. It was created using the TIES merge method from mergekit, building upon the mistralai/Mistral-Small-24B-Base-2501 as its base model. This merging approach aims to combine the distinct characteristics and strengths of several pre-trained models into a single, more versatile entity.

Merge Details

This model is a composite of three distinct language models:

  • TheDrummer/Cydonia-24B-v2: A model contributed by TheDrummer.
  • arcee-ai/Arcee-Blitz: Developed by arcee-ai.
  • cognitivecomputations/Dolphin3.0-Mistral-24B: From cognitivecomputations.

The TIES merge method was applied with specific density and weight parameters (0.5 for each constituent model) to integrate their knowledge effectively. The configuration also specified int8_mask: true and dtype: float16 for the merged output, indicating considerations for efficiency and precision.

Key Characteristics

As a TIES merge of specialized models, BuddyGlassKilledBonziBuddy is intended to inherit a broad range of capabilities from its diverse components. While specific performance benchmarks for the merged model are not provided, its foundation on Mistral-Small-24B-Base-2501 and the inclusion of models like Dolphin3.0 (known for instruction following) suggest a focus on robust general-purpose language understanding and generation, potentially excelling in areas where its constituent models demonstrated proficiency.