[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-peft-bench-fine-tuning-methods-benchmark-zh":3,"article-related-peft-bench-fine-tuning-methods-benchmark-zh":37,"series-research-d1c6850c-f832-471b-8beb-c0ebc809667d":88},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":19,"translated_content":10,"views":20,"is_premium":21,"created_at":22,"updated_at":22,"cover_image":11,"published_at":23,"rewrite_status":24,"rewrite_error":10,"rewritten_from_id":25,"slug":26,"category":27,"related_article_id":28,"status":29,"google_indexed_at":30,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":31,"topic_cluster_id":35,"embedding":36,"is_canonical_seed":21},"d1c6850c-f832-471b-8beb-c0ebc809667d","PEFT-Bench 讓微調比較更公平","\u003Cp data-speakable=\"summary\">PEFT-Bench 把 27 個 NLP 資料集與 7 種 PEFT 方法放進同一套流程，比的不只準確率，也把參數、速度和記憶體成本算進去。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>研究機構\u003C\u002Fstrong>：Brno University of Technology + Kempelen Institute of Intelligent Technologies\u003C\u002Fli>\u003Cli>\u003Cstrong>核心數據\u003C\u002Fstrong>：27 個 NLP 資料集\u003C\u002Fli>\u003Cli>\u003Cstrong>突破點\u003C\u002Fstrong>：PSCP 成本評分\u003C\u002Fli>\u003C\u002Ful>\u003Cp>對做大型語言模型的人來說，問題從來不只是「哪個微調方法分數最高」。更現實的是，哪個方法真的划算。算力、記憶體、訓練時間、推理速度，這些都會直接影響你能不能把方法帶進專案、產品，或是研究流程。\u003C\u002Fp>\u003Cp>這篇 PEFT-Bench 想解的，就是 PEFT 方法「不好公平比較」這件事。作者認為，現有評估太分散，常常只看少數任務，還常集中在非自回歸模型或傳統 NLU 基準。對現在大量使用的自回歸 \u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa> 來說，這樣的比較不夠完整，也不夠一致。\u003C\u002Fp>\u003Ch2>這篇論文要修的是哪個洞\u003C\u002Fh2>\u003Cp>PEFT，也就是參數高效率微調，存在的理由很直接：全量微調大型模型太貴。對很多團隊來說，不只是 \u003Ca href=\"\u002Ftag\u002Fgpu\">GPU\u003C\u002Fa> 成本高，還會碰到儲存、訓練時間和能源消耗的壓力。對學界或小團隊尤其明顯。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779179048497-jm5y.png\" alt=\"PEFT-Bench 讓微調比較更公平\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>但問題是，PEFT 方法雖然多，評估方式卻很碎。這篇摘要明講，過去很多工作不是只測 GLUE、SuperGLUE，就是資料與實驗細節不夠完整，讓別人很難重跑。結果就是，方法之間看起來像在比，但其實常常不是同一個起跑線。\u003C\u002Fp>\u003Cp>作者也點出可重現性問題。有些方法缺少開源實作，或是實驗描述不夠細，導致後續研究只能沿用別人的數字，而不是在同一設定下重做。對研究社群來說，這會讓比較失真；對開發者來說，則會讓選型更靠運氣。\u003C\u002Fp>\u003Ch2>PEFT-Bench 到底做了什麼\u003C\u002Fh2>\u003Cp>PEFT-Bench 的定位，是一套統一的端到端 benchmark。它不是只給一個分數，而是把資料集、任務、模型、方法與評估流程一起標準化，讓不同 PEFT 方法能在相同環境下比較。\u003C\u002Fp>\u003Cp>這個 benchmark 涵蓋 27 個資料集、12 種任務，分成三大類：自然語言理解與推理、數學、\u003Ca href=\"\u002Fnews\u002F8-ai-coding-assistants-for-enterprise-teams-zh\">程式\u003C\u002Fa>碼生成。NLU 部分再細分成 GLUE、SuperGLUE 和其他資料集。這個設計的重點，在於它不只看傳統分類任務，也把生成型任務拉進來，讓比較更接近現在 LLM 的實際使用情境。\u003C\u002Fp>\u003Cp>為了支撐這套流程，作者還做了 PEFT-Factory。這個框架建在 LLaMA-Factory 之上，並且對接 HuggingFace PEFT library 的現成方法。意思很簡單：不是每次都自己手工拼環境，而是希望新方法能更容易插進同一套評估管線裡。\u003C\u002Fp>\u003Cp>摘要沒有把 7 種方法完整列出來，但它明確說明，這些方法會在同一套流程下，針對各資料集與模型組合做訓練與評估。重點不是單次跑分，而是把比較條件盡量拉齊。\u003C\u002Fp>\u003Ch2>方法怎麼運作，白話講就是這樣\u003C\u002Fh2>\u003Cp>整個流程可以拆成三層：資料集與任務、語言模型與 PEFT 方法、\u003Ca href=\"\u002Fnews\u002Fconfident-ai-llm-evaluation-metrics-guide-zh\">評估指標\u003C\u002Fa>。先選一個方法，再選一個資料集，在同一個 instruction-fine-tuned 模型上做 supervised fine-tuning，最後把結果算成可比較的指標。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779179052187-2see.png\" alt=\"PEFT-Bench 讓微調比較更公平\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這裡有個細節很重要：作者把 instruction 透過資料集專屬模板放進每個樣本。這代表 benchmark 測的不是「裸模型」的適應能力，而是更貼近實務的指令式微調。對很多現在的 LLM 應用來說，這才是常態。\u003C\u002Fp>\u003Cp>除了看任務表現，PEFT-Bench 也看效率與穩定性。摘要提到會比較方法在有限資料下的表現，也包含穩定性實驗。換句話說，它不只問「能不能學會」，也問「學得穩不穩」。\u003C\u002Fp>\u003Cp>作者另外提出 PSCP，也就是 PEFT Soft Cost Penalties。這個分數把可訓練參數量、推理速度、訓練記憶體用量一起算進去。這是一個很實際的改動，因為很多方法在榜單上看起來漂亮，但一放到真實部署環境，成本就不漂亮了。\u003C\u002Fp>\u003Ch2>論文真正證明了什麼\u003C\u002Fh2>\u003Cp>這篇摘要傳達的重點，不是某個方法全面勝出，而是 trade-off 很明顯。根據提供的內容，LoRA 的表現較好；BitFit 和 LNTuning 則更有效率。這種結果其實很符合工程現場：你很少只看一個分數就決定採用，通常還得看你到底缺的是品質，還是資源。\u003C\u002Fp>\u003Cp>另一個重要訊號是，PEFT 方法雖然能學到任務結構，但在數學推理與程式碼生成上，可能會傷到 correctness。這點對開發者很關鍵，因為這類任務常常不是「大概對」就可以。少一個 \u003Ca href=\"\u002Ftag\u002Ftoken\">token\u003C\u002Fa>，答案、程式或證明就可能整個壞掉。\u003C\u002Fp>\u003Cp>摘要也提到 soft prompt 類方法比較難訓練。這不是說它們不能用，而是提醒你，方法的穩定性與調參難度，可能會影響實際導入成本。對研究人員來說，這會影響實驗效率；對產品團隊來說，則會影響上線風險。\u003C\u002Fp>\u003Cp>不過也要注意，這份摘要沒有公開完整 benchmark 數字。它沒有列出各任務的詳細分數、延遲、記憶體差異，也沒有把 7 種方法的完整清單全放出來。所以如果你想找的是精確排行榜，這份摘要還不夠。\u003C\u002Fp>\u003Ch2>對開發者的實際影響\u003C\u002Fh2>\u003Cp>這篇最直接的價值，是把「微調方法比較」從單一準確率，拉回到可部署性。對做內部模型、原型驗證，或是研究 baseline 的團隊來說，這很有用。因為真正要選方法時，你關心的不只是分數，還有訓練要吃多少顯存、推理會不會太慢、方法穩不穩。\u003C\u002Fp>\u003Cp>PSCP 的概念尤其適合這種決策。它把參數量、推理速度、訓練記憶體整合進同一個成本觀點，等於逼大家不要只看 accuracy。這對 GPU 緊、預算緊、部署條件緊的團隊，特別有感。\u003C\u002Fp>\u003Cp>另外，PEFT-Bench 也把評估面拉寬。它不只看傳統 NLU，還把數學與 code generation 放進來。這代表某個方法如果只是在舊基準上表現好，不一定能在更實際的生成任務裡站得住腳。對開發者來說，這種更廣的測試面，通常比單一榜單更有參考價值。\u003C\u002Fp>\u003Cp>不過，benchmark 再完整，也不能直接等於你的工作負載。你的資料分佈、提示詞格式、部署限制，都可能讓結果改變。這篇論文比較像是在幫你建立一個更公平的比較底座，而不是替你直接選出唯一答案。\u003C\u002Fp>\u003Ch2>限制與還沒回答完的問題\u003C\u002Fh2>\u003Cp>這份來源資料仍有幾個空白。首先，摘要沒有完整列出 7 種 PEFT 方法名稱，也沒有說明模型家族的更細節設定。其次，它沒有提供各任務的逐項結果，因此無法從摘要推回哪個方法在\u003Ca href=\"\u002Fnews\u002Fwhy-amazon-q-developer-is-wrong-future-coding-zh\">什麼\u003C\u002Fa>任務上最強。\u003C\u002Fp>\u003Cp>再來，雖然作者強調可重現性與公平比較，但 benchmark 本身還是有侷限。它可以改善比較環境，卻不能消除每個專案自己的差異。換到不同資料集、不同提示格式、不同服務條件，方法表現還是可能變。\u003C\u002Fp>\u003Cp>即便如此，PEFT-Bench 仍然是個重要方向。因為它處理的不是單一演算法，而是整個評估流程。對一個長期被「各自跑各自的」困擾的領域來說，先把比較規格統一起來，本身就是很有價值的進展。\u003C\u002Fp>\u003Cul>\u003Cli>PEFT-Bench 把 27 個資料集與 12 類任務放進同一套流程。\u003C\u002Fli>\u003Cli>它比較 7 種 PEFT 方法，並把效率與穩定性納入評估。\u003C\u002Fli>\u003Cli>PSCP 會把可訓練參數、推理速度、訓練記憶體一起算進成本。\u003C\u002Fli>\u003C\u002Ful>\u003Cp>總結來說，這篇不是在宣告某個新 adapter 贏了，而是在幫 PEFT 比較變得更誠實、更可重用，也更貼近部署現實。\u003C\u002Fp>","PEFT-Bench 把 27 個 NLP 資料集與 7 種 PEFT 方法放進同一套流程，比的不只準確率，也把參數、速度和記憶體成本算進去。","arxiv.org","https:\u002F\u002Farxiv.org\u002Fhtml\u002F2511.21285v3",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779179048497-jm5y.png",[13,14,15,16,17,18],"PEFT","fine-tuning","benchmark","LoRA","instruction tuning","LLM","zh",2,false,"2026-05-19T08:23:36.803043+00:00","2026-05-19T08:23:36.688+00:00","done","b7ad98f8-b186-45d1-8393-1ff330f16b14","peft-bench-fine-tuning-methods-benchmark-zh","research","4ed1af1c-05fe-425c-a296-464dbfca0e73","published","2026-05-19T09:00:32.157+00:00",[32,33,34],"PEFT-Bench 讓 27 個資料集、7 種方法在同一套規格下比較。","摘要顯示 LoRA 偏向表現，BitFit 與 LNTuning 偏向效率。","PSCP 把參數、速度與記憶體成本納入評分，適合看部署取捨。","0c35a120-52fc-41fc-afa3-d404eb934158","[-0.0225246,-0.011881989,0.0151073225,-0.08600404,-0.020183682,-0.02315604,-0.00076313654,0.01960462,0.016883627,0.009296294,0.0028477036,-0.03884015,0.026975043,-0.020605488,0.1088591,0.039764114,-0.010665755,0.0049237083,0.01799722,-0.026277412,0.01099159,0.031983037,-0.02849762,0.014754281,-0.007762111,-0.00074998924,0.018640723,0.0021268856,0.020329168,0.0055480287,0.010582588,0.011026561,0.019804372,0.004153034,0.009049863,-0.00479919,0.022654904,0.014243485,0.012260721,0.0012683923,-0.011058179,-0.014125032,0.006614867,-0.01438925,-0.039269537,0.0108386455,0.00034527018,-0.021462405,-0.013579327,0.008072062,0.018720755,0.017085157,0.0016016532,-0.15097255,0.0015321681,-0.019488377,0.0026139272,0.018955907,-0.0053967987,0.0067678597,-0.029607559,0.0005960629,-0.008202682,-0.0125131905,-0.012502811,-0.032258634,0.020360244,-0.016585814,-0.015260994,-0.011204481,-0.02260259,-0.018843986,0.015518524,-0.018615445,0.008152633,-0.023895804,0.00393258,-0.0048211645,0.008203645,0.012442584,0.008459562,-0.008470725,0.0022178383,-0.0065159528,-0.012185785,0.01520007,0.0022195731,0.0015691171,0.014462426,-0.005502985,-0.00040228068,0.008003235,-0.005391576,-0.015426116,0.0029402347,-0.00876327,-0.002197525,0.0050646155,-0.0033531284,-0.0018568394,-0.0030349141,-0.010805279,0.0062565347,0.011382472,0.005248498,0.009907419,0.0153775,-0.001990387,-0.011740674,0.017025355,-0.0033779843,-0.019340204,0.010774866,0.0010847523,0.011824911,-0.1456524,-0.02110461,0.01064706,-0.0152880335,-0.009350554,-0.016913598,0.006255169,-0.0058900234,0.043599393,-0.009187339,-0.0131669,0.025012841,0.004876117,-0.016930401,0.0041848035,-0.02003194,0.022610277,0.009184076,-0.0093574105,-0.01950019,0.019086346,0.023983879,-0.036288355,-0.026411476,-0.0052492134,-0.010171242,0.0075772246,-0.0045910683,-0.0031786019,-0.013992787,-0.014982574,-0.024540333,-0.0021863969,-0.013343138,0.0010969099,0.025199955,-0.0069751013,-0.015118358,-0.012172053,0.024889791,-0.008742226,0.017958988,0.0037876526,-0.012844472,0.041922763,0.008534771,-0.0037044093,-0.020915888,0.007823065,0.0057104207,0.034582656,0.0022344044,0.00073092786,0.013407395,0.029774342,0.010627964,-0.00401619,0.00299456,-0.008668131,0.0061408416,-0.014722141,0.0063099903,0.015710577,0.000188867,0.009735983,0.020176126,0.0028589908,0.004474277,0.0006817938,0.0030257062,-0.006529133,0.0012193017,0.0028692049,0.031979866,0.030392986,-0.034788493,0.0158549,0.03562188,-0.03128853,-0.008235183,-0.024579259,-0.0067845574,0.01987311,0.0076499996,0.02342408,0.005752715,-0.022251524,0.00922533,-0.000116570285,-0.010024223,-0.032132685,-0.008527771,-0.028063795,0.005103357,-0.0069040344,-0.018237546,-0.0064197723,-0.002996823,-0.0030728325,-0.013016217,-0.00663762,5.910138e-05,0.022697728,0.03167304,-0.003530941,-0.009524965,0.0016049658,0.008066864,-0.0018568634,-0.020923601,-0.016834354,-0.00263428,-0.012451806,-0.029877651,0.006495599,0.0041332142,0.015841497,0.01597874,-0.02607477,0.035651773,0.012155248,-0.023377996,0.015488376,0.0391855,0.017145157,-0.027000953,0.028784446,-0.00030796046,0.01638688,0.031464823,-0.013341333,0.02313994,-0.017293217,-0.012889058,-0.009283196,-0.013093897,0.003284473,-0.01212747,-0.010881212,-0.002554361,-0.0061761667,-0.0062702675,0.015855856,0.0028827682,-0.007357715,0.0009039306,0.0029374748,-0.009386413,-0.016818052,0.03011915,0.008265733,0.035027757,-0.018669639,-0.019465497,0.03503087,-0.036106274,0.0061850813,0.013262443,0.021005204,0.010513331,-0.033515032,-0.056419343,0.026841214,0.023127776,0.013639602,0.01623506,0.0077556847,0.0060504824,0.030678451,-0.009153056,0.013574142,-0.0062661553,-0.019690664,-0.009852094,0.004084833,-0.012853366,0.03476869,-0.026575722,-0.0043731374,-0.022929998,-0.024143199,0.009588458,-0.0036529687,-0.005527419,0.02295493,-0.0016077996,0.017599307,0.01183839,0.06092981,0.009965774,-0.013282025,-0.010657072,0.057123233,0.0053759604,-0.031397626,-0.010985155,-0.011479045,-0.0020488566,-0.022896884,0.005925853,-0.007932195,0.012302747,0.005829827,-0.0061856215,-0.026728347,-0.008274371,0.0022744243,-0.03440188,0.013390043,-0.019538134,0.0010285276,-0.009054196,-0.0075583993,0.021116626,-0.021633355,0.004988931,0.044711445,0.017251788,-0.03129629,-0.008914244,-0.00093082443,-0.026009185,-0.0096067,-0.041228954,-0.023595884,0.005732695,-0.0010600298,-0.0014436839,0.017623467,-0.01184327,0.03502424,0.0100737475,0.0021718177,-0.027222762,-0.008605227,0.028580276,0.0007990511,-0.0020304713,-0.026027584,-0.01692916,0.033266988,-0.01901264,-0.0075148246,0.031631317,-0.0025183323,-0.0029604174,-0.0014650297,-0.004696737,0.037874065,0.032164905,-0.028357752,-0.0038504167,-0.02244716,-0.026575856,-0.00092126935,0.01710526,0.023154264,-5.1945703e-05,0.0037568966,-0.005710734,0.010746182,-0.017653968,-0.01069781,0.019280968,0.005504017,0.009723281,0.019633662,-0.02097207,-0.022775387,0.005489583,-0.012033427,0.019236194,0.01358972,-0.009123075,-0.014026038,0.011394365,0.0069072125,0.014354404,0.015636899,-0.00869059,0.0077709453,-0.014030977,-0.0073615676,-0.0069471663,-0.027560893,-0.0011656584,0.009473032,-0.01304185,0.0072341254,0.0048325076,-0.0108358655,-0.016179806,0.033373043,-0.0037368792,-0.010580953,-0.0042802882,-0.0033502327,0.012132944,0.031157672,0.019886902,0.018561,-0.0062035257,0.014436187,0.018230682,0.016864734,-0.00969789,0.043759458,-0.00012647844,-0.008965823,-0.0024232874,0.004839695,-0.0068237684,-0.003047322,0.015146523,-0.02443374,0.0020501437,-0.017014006,-0.009984199,-0.012615706,-0.05532302,-0.0015126319,-0.03699888,0.024468323,-0.0030311227,-0.02714826,-0.009171284,-0.012656421,-0.0007222918,0.010417954,0.025326451,-0.019570963,0.009158311,-0.010359408,-0.013384593,-0.023258902,0.047992434,0.023494622,0.027239498,0.00420226,-0.0054760557,-0.020785334,-0.015187569,-0.029704817,-0.018718006,-0.0117122615,-0.0141166495,-0.031804997,-0.027404007,0.035654154,0.006709999,-0.0031504766,-0.013194582,-0.017361723,-0.036171142,0.004267407,-0.029426862,-0.0063151247,0.029283272,0.00071762403,0.0096346205,0.007025277,0.0069366912,0.020566968,0.023731055,-0.00019871138,0.0031014923,-0.017590389,-0.011279652,-0.022579925,0.0072271023,0.01159001,-0.008151068,0.018525418,-0.035620496,0.007972363,0.023541512,0.018366678,0.0042806757,-0.0021418363,-0.013121301,-0.00016522613,-0.028854884,-0.021424906,0.015395058,0.022364033,-0.02559021,0.015542954,0.023779593,-0.01591388,-0.010211021,-0.037122365,0.0066545885,0.021562232,0.0061419653,-0.009894481,-0.023549907,0.0047627753,0.02741611,0.023590643,-0.022691298,0.015787505,0.01111997,0.02151071,-0.010912327,-0.018913137,-0.03678059,-0.01345186,-0.02983535,0.027441487,-0.0148926,-0.021001745,0.0072882227,-0.008870586,0.01753855,0.025132837,0.0024295766,-0.003858513,-0.003121286,-0.0046039857,0.0069173924,-0.024506982,0.0069125574,0.023192665,0.007370822,-0.004660546,-0.005124687,0.0012066095,-0.007586698,-0.012860286,0.022986505,-0.08582473,-0.005940037,0.01935332,0.008519623,0.010349109,-0.0032609084,-0.007163838,-0.014629719,0.011902871,0.019536966,0.023529284,0.003221117,0.016075473,0.013973777,-0.0011666463,-0.024897981,-0.022109516,-0.034816947,0.030439425,-0.0068263286,0.0145628555,0.009027773,0.016363248,0.011764173,0.0072980174,-0.014366552,-0.0020158524,0.0068149995,0.0045017204,-0.012956694,-0.030978281,-0.008377866,-0.0036327003,0.024123147,-0.0017760196,-0.014651338,0.011215717,-0.02209821,0.011191322,-0.01469984,0.0101670595,0.015217843,-0.03650598,-0.020731337,-0.020881468,0.0046765204,-0.03048053,0.024654731,0.0034871765,0.009401671,-0.003532104,-0.012829793,-0.018795703,-0.021542398,-0.007007801,-0.013316097,-0.014875142,0.0009286629,-0.019497374,0.006826615,-0.01728337,-0.016648175,0.0063191834,0.025577074,-0.030476822,0.0067998134,0.0042792927,0.016189443,0.02204433,0.016411666,0.0031194217,-0.0324418,-0.0036704615,-0.0043831593,-0.010734485,0.009942372,0.0055605588,-0.021893578,-0.008119514,0.014937565,-0.048368566,-0.002682607,-0.07129296,-0.0069434894,-0.018249027,0.00550846,0.03886453,-0.022917354,0.014567881,-0.02972771,0.0051892,-0.0029436974,0.006805145,0.0026906973,0.000385454,-0.030272134,0.0041137,0.015626643,0.010778888,-0.0013185354,-0.011637822,-0.035987012,-0.0050710267,-0.046067294,-0.0095258355,-0.025462918,-0.018400216,0.0076380433,0.007499863,-0.0007427876,-0.007842703,-0.016386624,0.010591485,-0.15191597,-0.03175994,0.0029042305,0.007157501,0.0017807417,-0.0038444824,0.018923638,0.015299754,0.0030517117,0.0006740656,-0.011112806,-0.054719545,-0.0047528534,0.0151863545,0.009568135,0.087296106,-0.0016308916,0.0025710932,-0.008513597,-0.024167018,-0.0042538512,-0.02244484,-0.034751948,0.0051570106,0.0040133935,0.014127214,0.025288804,-0.031582326,-0.012359559,0.02180312,0.005083275,-0.021039141,-0.014678564,0.008956796,-0.008601418,-0.009730329,-0.016217226,-0.007358796,0.022857012,-0.0027298634,0.028550148,0.02150119,0.015176327,0.014686203,-0.027372088,-0.017979747,-0.01679938,-0.005264275,-0.0015623624,0.0036773495,-0.013248495,-0.058955014,0.019702794,-0.015248076,0.0067120916,0.011250044,-0.03739512,0.011518719,0.017693557,0.021582948,0.016136285,-0.012439318,0.012399954,0.015290669,-0.024156151,-0.0043528075,0.004703048,0.0067363014,0.0055745123,0.0023466211,0.0027755962,0.0043947767,0.0019067051,0.012626462,-0.0012454774,-0.028554806,0.025117867,0.009522842,0.03231695,-0.037071288,0.003937069,0.032166146,0.020493941,-0.012160253,0.01899802,0.012433495,0.0083823,0.012346654,0.006216024,-0.032493886,0.02636778,0.033511512,-0.01415996,-0.007167784,-0.00685783,0.01332995,-0.002496541,0.0139464,0.013091031,-0.00973758,-0.021988701,-0.019349786,0.015778936,-0.02100968,0.017894004,0.0056497105,0.005802317,0.016230274,0.05169902,0.0007618702]",{"tags":38,"relatedLang":47,"relatedPosts":51},[39,41,42,44,45],{"name":16,"slug":40},"lora",{"name":14,"slug":14},{"name":13,"slug":43},"peft",{"name":15,"slug":15},{"name":17,"slug":46},"instruction-tuning",{"id":28,"slug":48,"title":49,"language":50},"peft-bench-fine-tuning-methods-benchmark-en","PEFT-Bench compares fine-tuning methods fairly","en",[52,58,64,70,76,82],{"id":53,"slug":54,"title":55,"cover_image":56,"image_url":56,"created_at":57,"category":27},"e24e6e7a-6181-476b-8583-339d854cec68","confident-ai-llm-evaluation-metrics-guide-zh","Confident AI 的 LLM 評估指標指南","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779178456675-x5m6.png","2026-05-19T08:13:46.193772+00:00",{"id":59,"slug":60,"title":61,"cover_image":62,"image_url":62,"created_at":63,"category":27},"adfa9b15-68b6-44cc-b34d-ebcb02c31210","code-becomes-the-agent-harness-zh","程式碼成了代理引擎","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779173040130-zcyg.png","2026-05-19T06:43:29.625994+00:00",{"id":65,"slug":66,"title":67,"cover_image":68,"image_url":68,"created_at":69,"category":27},"eda7a80a-b234-4ada-90d1-a37b144251dc","rrfp-readiness-driven-pipeline-training-zh","RRFP 讓管線訓練跟著就緒跑","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779172442474-n21q.png","2026-05-19T06:33:31.287772+00:00",{"id":71,"slug":72,"title":73,"cover_image":74,"image_url":74,"created_at":75,"category":27},"475844e6-3e2c-49a6-aea0-86a94945d2c2","dashattention-differentiable-adaptive-sparse-attention-zh","DashAttention 讓稀疏長上下文可微","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779171840613-dq1r.png","2026-05-19T06:23:32.886786+00:00",{"id":77,"slug":78,"title":79,"cover_image":80,"image_url":80,"created_at":81,"category":27},"23a3d4c7-5cb7-40ae-a05b-1542364e786f","ibm-prompt-guide-turns-ai-guesses-into-outputs-zh","IBM 提示指南把猜答案變輸出","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779132863293-etob.png","2026-05-18T19:33:55.711767+00:00",{"id":83,"slug":84,"title":85,"cover_image":86,"image_url":86,"created_at":87,"category":27},"7c89c3bd-48cb-4b4e-942d-bbf0409fc392","cattle-trade-llm-bluffing-bargaining-benchmark-zh","Cattle Trade 要測 LLM 談判 bluffing","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779085437419-b0zw.png","2026-05-18T06:23:27.885037+00:00",[89,94,99,104,109,114,119,124,129,134],{"id":90,"slug":91,"title":92,"created_at":93},"f18dbadb-8c59-4723-84a4-6ad22746c77a","deepmind-bets-on-continuous-learning-ai-2026-zh","DeepMind 押注 2026 連續學習 AI","2026-03-26T08:16:02.367355+00:00",{"id":95,"slug":96,"title":97,"created_at":98},"f4a106cb-02a6-4508-8f39-9720a0a93cee","ml-papers-of-the-week-github-research-desk-zh","每週 ML 論文清單，為何紅到 GitHub","2026-03-27T01:11:39.284175+00:00",{"id":100,"slug":101,"title":102,"created_at":103},"c4f807ca-4e5f-47f1-a48c-961cf3fc44dc","ai-ml-conferences-to-watch-in-2026-zh","2026 AI 研討會投稿時程整理","2026-03-27T01:51:53.874432+00:00",{"id":105,"slug":106,"title":107,"created_at":108},"9f50561b-aebd-46ba-94a8-363198aa7091","openclaw-agents-manipulated-self-sabotage-zh","OpenClaw Agent 會自己搞砸自己","2026-03-28T03:03:18.786425+00:00",{"id":110,"slug":111,"title":112,"created_at":113},"11f22e92-7066-4978-a544-31f5f2156ec6","vega-learning-to-drive-with-natural-language-instructions-zh","Vega：使用自然語言指示進行自駕車控制","2026-03-28T14:54:04.847912+00:00",{"id":115,"slug":116,"title":117,"created_at":118},"a4c7cfec-8d0e-4fec-93cf-1b9699a530b8","drive-my-way-en-zh","Drive My Way：個性化自駕車風格的實現","2026-03-28T14:54:26.207495+00:00",{"id":120,"slug":121,"title":122,"created_at":123},"dec02f89-fd39-41ba-8e4d-11ede93a536d","training-knowledge-bases-with-writeback-rag-zh","用 WriteBack-RAG 強化知識庫提升檢索效能","2026-03-28T14:54:45.775606+00:00",{"id":125,"slug":126,"title":127,"created_at":128},"3886be5c-a137-40cc-b9e2-0bf18430c002","packforcing-efficient-long-video-generation-method-zh","PackForcing：短影片訓練也能生成長影片","2026-03-28T14:55:02.688141+00:00",{"id":130,"slug":131,"title":132,"created_at":133},"72b90667-d930-4cc9-8ced-aaa0f8968d44","pixelsmile-toward-fine-grained-facial-expression-editing-zh","PixelSmile：提升精細臉部表情編輯的新方法","2026-03-28T14:55:20.678181+00:00",{"id":135,"slug":136,"title":137,"created_at":138},"cf046742-efb2-4753-aef9-caed5da5e32e","adaptive-block-scaled-data-types-zh","IF4：神經網路量化的聰明選擇","2026-03-31T06:00:36.990273+00:00"]