uukuguy/Mistral-7B-OpenOrca-lora-merged

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kLicense:llama2Architecture:Transformer0.0K Open Weights Cold

The uukuguy/Mistral-7B-OpenOrca-lora-merged model is a regenerated 7 billion parameter language model combining the Mistral-7B-v0.1 base with a LoRA adapter extracted from the Mistral-7B-OpenOrca model. This model is designed to verify if a merged LoRA can achieve comparable performance to the original fine-tuned model. It aims to serve as a component in a toolkit for dynamically loading and switching multiple LoRA modules based on user queries.

Loading preview...

Overview

This model, uukuguy/Mistral-7B-OpenOrca-lora-merged, is a 7 billion parameter language model created by merging the base model Mistral-7B-v0.1 with a LoRA (Low-Rank Adaptation) module. The LoRA module was specifically extracted from the Mistral-7B-OpenOrca model, which is known for its efficient parameter fine-tuning. The primary goal behind this merged model is to validate whether a LoRA-merged model can replicate the performance of its original fine-tuned counterpart.

Key Capabilities & Purpose

  • LoRA Verification: The model's core purpose is to test the efficacy of LoRA extraction and merging, aiming to achieve performance comparable to the original Mistral-7B-OpenOrca model.
  • Foundation for Multi-LoRA Systems: It serves as a foundational step towards developing a toolkit capable of loading and dynamically switching multiple LoRA modules based on user queries, optimizing response generation.

Performance Insights

Evaluations show that the r=256 configuration of the LoRA-merged model achieves competitive scores, sometimes surpassing the original Mistral-7B-OpenOrca in specific benchmarks:

  • MMLU_acc (5-shot): 64.28 (vs. 61.42 for original)
  • HellaSwag_acc_norm (10-shot): 84 (vs. 83 for original)
  • Open LLM Score: 65.81 (vs. 65.11 for original)

Training Details

The model was trained using bitsandbytes quantization with load_in_4bit: True, bnb_4bit_quant_type: nf4, and bnb_4bit_use_double_quant: True, leveraging PEFT 0.5.0.