Course Code: sbmftm
Duration: 14 hours
Prerequisites:
  • 了解機器學習模型與訓練流程
  • 具備微調與LLMs的實務經驗
  • 熟悉Python與NLP概念

目標受眾

  • AI合規團隊
  • ML工程師
Overview:

Safety and Bias Mitigation in Fine-Tuned Models 隨著人工智慧在各行業的決策中日益深入，以及監管標準的不斷演進，已成為一個日益關注的問題。

這項由講師指導的培訓（線上或線下）針對中級機器學習工程師和人工智慧合規專業人員，旨在幫助他們識別、評估並減少微調語言模型中的安全風險和偏見。

培訓結束後，參與者將能夠：

  • 了解安全人工智慧系統的倫理和監管背景。
  • 識別並評估微調模型中的常見偏見形式。
  • 在訓練期間及之後應用偏見緩解技術。
  • 設計並審核模型，確保其安全性、透明性和公平性。

課程形式

  • 互動式講座與討論。
  • 大量練習與實踐。
  • 在即時實驗室環境中進行實作。

課程定制選項

  • 如需為本課程定制培訓，請聯繫我們安排。
Course Outline:

安全與公平AI的基礎

  • 關鍵概念：安全性、偏見、公平性、透明度
  • 偏見類型：數據集偏見、代表性偏見、算法偏見
  • 監管框架概述（歐盟AI法案、GDPR等）

微調模型中的偏見

  • 微調如何引入或放大偏見
  • 案例研究與現實中的失敗案例
  • 識別數據集和模型預測中的偏見

偏見緩解技術

  • 數據層面策略（重新平衡、數據增強）
  • 訓練中策略（正則化、對抗性去偏見）
  • 後處理策略（輸出過濾、校準）

模型安全與穩健性

  • 檢測不安全或有害的輸出
  • 處理對抗性輸入
  • 紅隊演練與壓力測試微調模型

審計與監控AI系統

  • 偏見與公平性評估指標（如人口統計平等）
  • 可解釋性工具與透明度框架
  • 持續監控與治理實踐

工具包與實踐操作

  • 使用開源庫（如Fairlearn、Transformers、CheckList）
  • 實踐操作：檢測與緩解微調模型中的偏見
  • 通過提示設計與約束生成安全輸出

企業Use Case與合規準備

  • 在LLM工作流程中整合安全性的最佳實踐
  • Documentation與模型卡片用於合規
  • 準備審計與外部審查

總結與下一步

Sites Published:

United Arab Emirates - Safety and Bias Mitigation in Fine-Tuned Models

Qatar - Safety and Bias Mitigation in Fine-Tuned Models

Egypt - Safety and Bias Mitigation in Fine-Tuned Models

Saudi Arabia - Safety and Bias Mitigation in Fine-Tuned Models

South Africa - Safety and Bias Mitigation in Fine-Tuned Models

Brasil - Safety and Bias Mitigation in Fine-Tuned Models

Canada - Safety and Bias Mitigation in Fine-Tuned Models

中国 - Safety and Bias Mitigation in Fine-Tuned Models

香港 - Safety and Bias Mitigation in Fine-Tuned Models

澳門 - Safety and Bias Mitigation in Fine-Tuned Models

台灣 - Safety and Bias Mitigation in Fine-Tuned Models

USA - Safety and Bias Mitigation in Fine-Tuned Models

Österreich - Safety and Bias Mitigation in Fine-Tuned Models

Schweiz - Safety and Bias Mitigation in Fine-Tuned Models

Deutschland - Safety and Bias Mitigation in Fine-Tuned Models

Czech Republic - Safety and Bias Mitigation in Fine-Tuned Models

Denmark - Safety and Bias Mitigation in Fine-Tuned Models

Estonia - Safety and Bias Mitigation in Fine-Tuned Models

Finland - Safety and Bias Mitigation in Fine-Tuned Models

Greece - Safety and Bias Mitigation in Fine-Tuned Models

Magyarország - Safety and Bias Mitigation in Fine-Tuned Models

Ireland - Safety and Bias Mitigation in Fine-Tuned Models

Luxembourg - Safety and Bias Mitigation in Fine-Tuned Models

Latvia - Safety and Bias Mitigation in Fine-Tuned Models

España - Safety and Bias Mitigation in Fine-Tuned Models

Italia - Safety and Bias Mitigation in Fine-Tuned Models

Lithuania - Safety and Bias Mitigation in Fine-Tuned Models

Nederland - Safety and Bias Mitigation in Fine-Tuned Models

Norway - Safety and Bias Mitigation in Fine-Tuned Models

Portugal - Safety and Bias Mitigation in Fine-Tuned Models

România - Safety and Bias Mitigation in Fine-Tuned Models

Sverige - Safety and Bias Mitigation in Fine-Tuned Models

Türkiye - Safety and Bias Mitigation in Fine-Tuned Models

Malta - Safety and Bias Mitigation in Fine-Tuned Models

Belgique - Safety and Bias Mitigation in Fine-Tuned Models

France - Safety and Bias Mitigation in Fine-Tuned Models

日本 - Safety and Bias Mitigation in Fine-Tuned Models

Australia - Safety and Bias Mitigation in Fine-Tuned Models

Malaysia - Safety and Bias Mitigation in Fine-Tuned Models

New Zealand - Safety and Bias Mitigation in Fine-Tuned Models

Philippines - Safety and Bias Mitigation in Fine-Tuned Models

Singapore - Safety and Bias Mitigation in Fine-Tuned Models

Thailand - Safety and Bias Mitigation in Fine-Tuned Models

Vietnam - Safety and Bias Mitigation in Fine-Tuned Models

India - Safety and Bias Mitigation in Fine-Tuned Models

Argentina - Safety and Bias Mitigation in Fine-Tuned Models

Chile - Safety and Bias Mitigation in Fine-Tuned Models

Costa Rica - Safety and Bias Mitigation in Fine-Tuned Models

Ecuador - Safety and Bias Mitigation in Fine-Tuned Models

Guatemala - Safety and Bias Mitigation in Fine-Tuned Models

Colombia - Safety and Bias Mitigation in Fine-Tuned Models

México - Safety and Bias Mitigation in Fine-Tuned Models

Panama - Safety and Bias Mitigation in Fine-Tuned Models

Peru - Safety and Bias Mitigation in Fine-Tuned Models

Uruguay - Safety and Bias Mitigation in Fine-Tuned Models

Venezuela - Safety and Bias Mitigation in Fine-Tuned Models

Polska - Safety and Bias Mitigation in Fine-Tuned Models

United Kingdom - Safety and Bias Mitigation in Fine-Tuned Models

South Korea - Safety and Bias Mitigation in Fine-Tuned Models

Pakistan - Safety and Bias Mitigation in Fine-Tuned Models

Sri Lanka - Safety and Bias Mitigation in Fine-Tuned Models

Bulgaria - Safety and Bias Mitigation in Fine-Tuned Models

Bolivia - Safety and Bias Mitigation in Fine-Tuned Models

Indonesia - Safety and Bias Mitigation in Fine-Tuned Models

Kazakhstan - Safety and Bias Mitigation in Fine-Tuned Models

Moldova - Safety and Bias Mitigation in Fine-Tuned Models

Morocco - Safety and Bias Mitigation in Fine-Tuned Models

Tunisia - Safety and Bias Mitigation in Fine-Tuned Models

Kuwait - Safety and Bias Mitigation in Fine-Tuned Models

Oman - Safety and Bias Mitigation in Fine-Tuned Models

Slovakia - Safety and Bias Mitigation in Fine-Tuned Models

Kenya - Safety and Bias Mitigation in Fine-Tuned Models

Nigeria - Safety and Bias Mitigation in Fine-Tuned Models

Botswana - Safety and Bias Mitigation in Fine-Tuned Models

Slovenia - Safety and Bias Mitigation in Fine-Tuned Models

Croatia - Safety and Bias Mitigation in Fine-Tuned Models

Serbia - Safety and Bias Mitigation in Fine-Tuned Models

Bhutan - Safety and Bias Mitigation in Fine-Tuned Models

Nepal - Safety and Bias Mitigation in Fine-Tuned Models

Uzbekistan - Safety and Bias Mitigation in Fine-Tuned Models