digitous/13B-Chimera

TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kPublished:May 23, 2023Architecture:Transformer0.0K Cold

digitous/13B-Chimera is an experimental 13 billion parameter language model created by digitous, resulting from a unique composite merge of several models and LoRAs. It aims to additively combine features from components like MantiCore3E, VicunaCocktail, SuperCOT, StorytellingV2, SuperHOT Prototype (8192 context), and Metharme. This model is designed to explore the synergistic application of desired features without diluting effective behavior, showing promising subjective results in its experimental application.

Loading preview...

13B-Chimera Overview

digitous/13B-Chimera is an experimental 13 billion parameter language model developed by digitous, built through a novel composite merging technique. This model combines several base models and LoRAs, including MantiCore3E, VicunaCocktail, SuperCOT, StorytellingV2, SuperHOT Prototype (with an 8192 context length), and Metharme. The core idea behind Chimera is to apply desired features additively, aiming to enhance capabilities without diluting the model's overall effective behavior.

Key Characteristics

  • Experimental Merging: Utilizes an unconventional method of applying LoRAs and model merges on non-native LLaMA base models.
  • Feature Composition: Integrates components known for diverse strengths, such as reasoning (SuperCOT), storytelling (StorytellingV2), and extended context (SuperHOT Prototype).
  • Subjective Promise: Initial subjective evaluations indicate very promising results, though further objective testing is required.

Good for

  • Experimental AI Development: Ideal for researchers and developers interested in exploring advanced model merging and LoRA application techniques.
  • Creative Text Generation: Components like StorytellingV2 and Metharme suggest potential for enhanced narrative and conversational outputs.
  • Extended Context Applications: The inclusion of SuperHOT Prototype with 8192 context length makes it suitable for tasks requiring longer memory or input processing.
  • Custom Instruction Following: Verified to work with Alpaca and Vicuna instruct formats, offering flexibility for various instruction-tuned tasks.