mannix / phi3-mini-cpo-simpo

Phi-3-mini-4K-instruct with CPO-SimPO

3B

30 Pulls Updated 2 months ago

2 months ago

d1e679dd05f7 · 2.9GB

model
phi3
·
3.82B
·
Q5_1
license
Microsoft. Copyright (c) Microsoft Corporation. MIT License Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED *AS IS*, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
params
{"stop":["<|end|>","<|user|>","<|assistant|>"]}
template
{{ if .System }}<|system|> {{ .System }}<|end|> {{ end }}{{ if .Prompt }}<|user|> {{ .Prompt }}<|end|> {{ end }}<|assistant|> {{ .Response }}<|end|>

Readme

Phi-3-mini-4K-instruct with CPO-SimPO

  • Quantizations with i-matrix calibration_datav3.txt
  • Saftensors converted to fp32

This repository contains the Phi-3-mini-4K-instruct model enhanced with the CPO-SimPO technique. CPO-SimPO combines Contrastive Preference Optimization (CPO) and Simple Preference Optimization (SimPO).

Introduction

Phi-3-mini-4K-instruct is a model optimized for instruction-based tasks. This approach has demonstrated notable improvements in key benchmarks, pushing the boundaries of AI preference learning.

What is CPO-SimPO?

CPO-SimPO is a novel technique, which combines elements from CPO and SimPO:

  • Contrastive Preference Optimization (CPO): Adds a behavior cloning regularizer to ensure the model remains close to the preferred data distribution.
  • Simple Preference Optimization (SimPO): Incorporates length normalization and target reward margins to prevent the generation of long but low-quality sequences.

Github

CPO-SIMPO

Model Performance

COMING SOON!

Key Improvements:

  • Enhanced Model Performance: Significant score improvements, particularly in GSM8K (up by 8.49 points!) and TruthfulQA (up by 2.07 points).
  • Quality Control: Improved generation of high-quality sequences through length normalization and reward margins.
  • Balanced Optimization: The BC regularizer helps maintain the integrity of learned preferences without deviating from the preferred data distribution.