【中文版】Claude3技術(shù)報(bào)告-Claude3模型系列作品、十四行詩(shī)、俳句

22.1預(yù)期用途2.2意外用途2.3禁止用途們的產(chǎn)品盡可能安全和值得信賴。2.4防止濫用3456一般推理數(shù)學(xué)解決問(wèn)題十四行詩(shī)40.5%43.1%40.9%42.5%小學(xué)數(shù)學(xué)人類評(píng)估Python編碼任務(wù)多語(yǔ)言數(shù)學(xué)算術(shù)95.0%90.7%SFT、5次CoT67.0%648.1%74.4%）（74.5%779.0%少校1@32評(píng)價(jià)不一海拉斯瓦格生物醫(yī)學(xué)問(wèn)題Python編碼任務(wù)代碼生成5發(fā)75.8%0射門74.9%5發(fā)92.9%0射門率70.2%通過(guò)@186.4%78.3%79.7%83.1%756786787分子束外延––––––––8→藝術(shù)與設(shè)計(jì)→商業(yè)→科學(xué)→健康與醫(yī)藥→人文社會(huì)科學(xué)→技術(shù)與工程全面的文檔理解MathVista[54](testmini)數(shù)學(xué)科學(xué)圖表克勞德3作品67.5%67.2%48.9%61.1%70.0%50.6%59.4%50.5%?80.8%?克勞德3克勞德3十四行詩(shī)俳句61.7%60.8%58.2%52.5%37.1%37.1%57.1%52.3%68.7%66.0%45.0%41.5%53.1%50.2%47.9%?46.4%?88.7%86.7%81.1%?81.7%?GPT?4V1165.8%59.3%54.7%64.7%72.5%36.7%56.8%（來(lái)自[3]）49.9%78.2%78.5%?雙子座70.0%56.7%48.0%67.3%78.3%47.1%59.4%90.9%53%79.5%雙子座58.5%52.1%雙子座47.9%45.2%73.9%74.1%95.4.1拒絕R=圖5該圖顯示了每個(gè)任務(wù)的人類偏好勝率與基線ClaudeInstant?？藙诘?作品8發(fā)90.5%0射門率90.7%克勞德383.7%83.5%克勞德3俳句76.5%75.1%GPT?4374.5%79%雙子座超4雙子座88.7%雙子座63.5%作品俳句克洛德5發(fā)79.1% 85.9%84.9%80.2%79.4%克勞德2.1克勞德2.0克勞德2.1克勞德2.085.5%82.8%84.3%85.5%82.8%84.3%80.5%78.7%99.4%95.4%91.4%95.9%91.9%94.5%92.7%文件不包含有關(guān)披薩配料的任何其他信息。6.1負(fù)責(zé)任的擴(kuò)展政策?作品俳句GPT?413GPT?3.513鉆石擴(kuò)展放50.4%53.3%49.1%50.2%48.8%49.2%40.4%42.9%46.3%38.5%39.1%38.0%39.3%33.3%36.4%40.1%35.2%36.4%34.7%35.5%35.7%38.8%–39.5%39.7%38.1%38.7%28.1%29.6%–28.9%28.0%28.4%28.2%擇的解和推理基準(zhǔn)通用人工智能?！?023年。com/files/4zrzovbb/website/bd2a28d2535bfb0494cc8e2a3bf135d2e7523226.pdf。/index/anthropics?responsible?scaling?policy。/index/claudes?constitution。2019。http://papers.neurips.cc/paper/9015?pytorch?an?imperative?style?high?performance?deep?learning?library.pdf。/index/evaluating?ai?systems。得到教訓(xùn)?！?022年8月。/index/red?teaming?language?models?to?reduce?harms?methods?scaling?behaviors?and?lessons?learned。[12]Anthropic，“大型語(yǔ)言模型中的道德自我糾正能力?！?023年2月。/index/the?capacity?for?moral?self?Correction?in?large?language?models。表示。”2023年。/index/frontier?threats?red?teaming?for?ai?safety。index/collective?constitutional?ai?aligning?a?language?model?with?public?input。[18]“HH?RLHF數(shù)據(jù)集卡”，https://huggingface.co/datasets/Anthropic/hhnistpubs/ai/NIST.AI.100?1.pdf。[21]“人擇隱私政策”。2023年7月。/legal/力?！?023年6月。技術(shù)報(bào)告?！?023年。/deepmind?media/gemini/gemini_v15report.pdf。挑戰(zhàn)問(wèn)題。”2023年4月。/news?resources/ncbe?releases?first?full?length?simulated?mbe?study?aid，2021年。(2016)。https://api.sem/papers/GPTV_System_Card.p/internet/2024/01/31/americans?social?media?use，2024/briefing?room/statements?releases/2023/07/21/fact?ssecures?voluntory?commitments?from?le/briefing?room/statements?releases/2023/10/30/fact?sheet?president?biden?issues?executive?order?on?safe?secure?and?值得信賴的人工智能