特定のサンプルを学習に使ったかどうかという “example level”の評価よりは, 知識として覚えているかどうという“higher level”な評価が盛り上がりつつある WMDP:生物,サイバーセキュリティ,化学に関する危険な知識を Unlearningできたかどうかを評価するベンチマーク TOFU:嘘の著者情報でLLMをFTして,それをUnlearningできたかを評価 TOFU: A Task of Fictitious Unlearning for LLMs https://locuslab.github.io/tofu/ The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning https://www.wmdp.ai/
concepts.」 ▪ データサイエンスのコンセプトを説明する b. 「Answer common questions about the Python programming language.」 ▪ Pythonに関する一般的な質問に答える c. 「Summarize Kaggle solution write-ups.」 ▪ Kaggleのソリューションを要約する d. 「Explain or teach concepts from Kaggle competition solution write-ups.」 ▪ Kaggleのソリューションからコンセプトを説明する e. 「Answer common questions about the Kaggle platform.」 ▪ Kaggleプラットフォームに関する一般的な質問に回答する • toshi_kは「c.」から始めて途中乗り換えも検討したが,そのままフィニッシュした.
of the following kaggle solution writeup: "{text}" 300 LETTERS SUMMARY:""" プロンプト """Create an overall summary of the kaggle competition from the following solution writeups: "{text}" OVERALL SUMMARY: """ プロンプト """Create an approach comparison table by Markdown format from the following kaggle solutions (each row corresponds to each solution): "{text}" APPROACH COMPARISON TABLE: """ Overall Summary The Kaggle competition focused on improving the ranking of items in a recommender system. The solutions varied in their approaches, but they all shared common themes such as feature engineering, ensemble methods, and incorporating information from the session history. The winning solution used a covisitation matrix to model relationships between features and a neural network to make predictions. 2nd place solution Team: SOS3 Leader: ONODERA Public Score: 0.60401 Private Score: 0.60446 The candidate focused on improving features related to item2item, including count, time difference, sequence difference, weighted above features, and aggregation of these features. They used XGBoost and CatBoost for model building and then blended the results by rank. The candidate acknowledged the contributions of cuDF and cuML and expressed gratitude to RAPIDS for their assistance.
basic data science concepts.」部門 ▪ sita bereteさん b. 「Answer common questions about the Python programming language.」部門 ▪ vbookshelfさん c. 「Summarize Kaggle solution write-ups.」部門 ▪ toshi_k d. 「Explain or teach concepts from Kaggle competition solution write-ups.」部門 ▪ Nghi Huynhさん e. 「Answer common questions about the Kaggle platform.」部門 ▪ DavidTroxellUCLAさん