Nitral-Archive/Eris-Daturamix-7b

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kPublished:Mar 8, 2024License:otherArchitecture:Transformer0.0K Cold

Eris-Daturamix-7b is a 7 billion parameter language model created by Nitral-Archive, resulting from a merge of Test157t/Eris-Floramix-7b and ResplendentAI/Datura_7B. This model leverages a slerp merge method with specific parameter weighting for self_attn and mlp layers. It is designed for general language tasks, combining the strengths of its constituent models.

Loading preview...

Eris-Daturamix-7b: Merged Language Model

Eris-Daturamix-7b is a 7 billion parameter language model developed by Nitral-Archive, created through a strategic merge of two distinct base models: Test157t/Eris-Floramix-7b and ResplendentAI/Datura_7B.

Merge Configuration

The model was constructed using a slerp (spherical linear interpolation) merge method. This process involved combining the full 32 layers of both source models. A key aspect of its configuration is the differentiated weighting applied to specific architectural components:

  • Self-attention layers (self_attn): These layers were merged with varying t values, specifically [0, 0.5, 0.3, 0.7, 1] across different filters.
  • MLP layers (mlp): These layers also received distinct t values, specifically [1, 0.5, 0.7, 0.3, 0] across different filters.
  • General parameters: A default t value of 0.5 was applied to other parameters.

This precise merging strategy aims to integrate the unique characteristics and capabilities of both Eris-Floramix-7b and Datura_7B into a single, cohesive model. The model was produced with bfloat16 dtype.

Good for

  • General language generation tasks
  • Exploring merged model performance