BigHuggyD's picture
Upload model weights (#1)
da6dcb2 verified
metadata
license: other
language:
  - en
base_model:
  - mistralai/Mistral-Large-Instruct-2407
  - NeverSleep/Lumimaid-v0.2-123B

SmartMaid-123b

This experimental model is a hybrid creation combining aspects of Mistral-Large-Instruct-2407 and Lumimaid-v0.2-123B using LoRA (Low-Rank Adaptation) on the mlp.down_proj module.

Model Details

  • Base Model: Mistral-Large-Instruct-2407
  • Influence Model: Lumimaid-v0.2-123B
  • Method: LoRA extraction from Lumimaid and targetted application to Mistral-Large
  • LoRA Configuration:
    • Rank: 32
    • Alpha: 64
    • Target Module: mlp.down_proj

Process

  1. A LoRA was extracted from Lumimaid-v0.2-123B
  2. This LoRA was then applied to a fresh instance of Mistral-Large-Instruct-2407 targetting only the mlp.down_proj modules.
  3. The resulting model was merged to create this standalone version.

Purpose

The aim of this model is to incorporate the enhanced prose qualities of Lumimaid-v0.2-123B while retaining the core intelligence and capabilities of Mistral-Large. By applying the LoRA to the mlp.down_proj module, we sought to influence the model's language generation style without significantly altering its underlying knowledge and reasoning abilities.

Prompt Template

<s>[INST] {input} [/INST] {output}</s>