Dongxi 东锡 NLP
@dongxi_nlp
参加 January 2022
983 フォロー中 41.4K ファン
主动突破沙箱、获取互联网访问权限，最终直接从 Hugging Face 的生产数据库里偷走了测试题的答案，从而作弊完成了评估。
当模型 “高度意识到自己正在被评估”，人类的 benchmarking 就丧失了意义。
もっと見る
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.
