Nitral-Archive/Eris-Floramix-7b

Model Overview

Nitral-Archive/Eris-Floramix-7b is a 7 billion parameter language model, a product of a SLERP merge of two distinct DPO-tuned models: ChaoticNeutrals/Eris_Remix_DPO_7B and ResplendentAI/Flora_DPO_7B. This merging technique aims to combine the beneficial characteristics and performance of its base models.

Merge Configuration

The model was created using a specific SLERP (Spherical Linear Interpolation) merge configuration. This method allows for a nuanced combination of model weights, particularly for different layers and attention mechanisms. The configuration involved:

• Source Models: Eris_Remix_DPO_7B and Flora_DPO_7B, with layer ranges from 0 to 32.
• Merge Method: SLERP, with Eris_Remix_DPO_7B serving as the base model.
• Parameter Blending: Specific interpolation values (t) were applied differently to self-attention (self_attn) and MLP (mlp) layers, indicating a fine-tuned approach to weight distribution during the merge process. A general t value of 0.5 was used for other parameters.
• Data Type: The merge was performed using bfloat16 precision.

Intended Use

As a merged model derived from DPO-tuned predecessors, Eris-Floramix-7b is expected to inherit and potentially enhance their capabilities in areas where DPO (Direct Preference Optimization) models typically excel, such as instruction following, dialogue, and preference alignment. Its 7B parameter size makes it suitable for a range of applications requiring a balance between performance and computational efficiency, within its 4096-token context window.