InoteveraGaidhi rinotevera
Kurumidza Kushandura uye Regression Testing
Tekinoroji
Nhungamiro yehunyanzvi
Otomatiki nekukurumidza optimization inoshandisa algorithm kutsvaga zvirinani mirairo kana mashoma-epfuti mienzaniso.
Teerera kugwaro iriAI-yakarondedzerwa · ~3 min
Inopa zvibodzwa zvemumiriri zvinopesana nemetric pane seti yemienzaniso pane kuvimba nemaoko-tweaked mazwi. Maitiro anozivikanwa anosanganisira APE, Google DeepMind's OPRO uye DSPy, yakavhurika-sosi chimiro kubva kuStanford. Pamwe chete vanoshandura huinjiniya nekukurumidza kubva pakufungidzira kuita inoyerwa, inodzokororwa maitiro.
Kurudziro yemaoko ine kushaya simba kunozivikanwa. Izwi rinoshanda kune imwe modhi, dhatabheti kana shanduro inogona kunyarara kushanda pane imwe, uye shanduko dzinowanzo kutongwa pamienzaniso mishoma. Otomatiki nekukurumidza optimization inotsiva icho loop nekutsvaga. Unopa zvinhu zvitatu: - Basa. - Seti yezvipo, zvine mutsindo nezvinotarisirwa zvinobuda. - Metric, senge mutambo chaiwo kana rubriki inopihwa neimwe modhi. Iyo optimizer inogadzira mafungidziro evamiriri, anoamhanyisa, anoaisa zvibodzwa uye anochengeta vakundi. Mienzaniso yepakutanga yakaita kuti pfungwa yacho ive yakasimba. APE (Automatic Prompt Engineer, Zhou et al., 2022) vaive nemuenzaniso waikurudzira mirairo yakawanda kubva mumienzaniso yekupinza-yekubuda, ndokuzochengeta iwo epamusoro-zvibodzwa. OPRO (Optimization by Prompting, Yang et al., Google DeepMind, 2023) yakashandisa LLM se optimizer. Yayo meta-yekukurumidza yakanyorwa mirairo yekutanga nezvibodzwa zvavo, uye modhi yakakurudzira mitsva. OPRO yakawana mirairo yakadai sekuti "femera zvakadzama uye shanda padambudziko iri nhanho-nhanho," iyo yakawana zvakanaka kune imwe modhi pagiredhi-chikoro math. Mazwi akakwenenzverwa anogona kuve asina kujeka uye akananga kune modhi. DSPy inoenderera mberi nekuchinja kuti zvirongwa zvinonyorwa sei. Yakagadzirwa kuStanford NLP naOmar Khattab nevabatsiri uye yakaunzwa muna 2023. Iwe unozivisa zvinoitwa nedanho rega rega nesaini, se "mubvunzo, mamiriro -> mhinduro," uye batanidza mamodule se Predict kana ChainOfThought. An optimizer (yakambodaidzwa kuti teleprompter), yakadai seBootstrapFewShot kana MIPROv2, uye "inounganidza" chirongwa nekusarudza zviratidzo uye mirairo yemodule yega yega kuti uwedzere metric yako. Mafungiro matatu asiri iwo akakodzera kugadziriswa: - Maturusi aya anobvisa kutonga kwevanhu. Ivo havadaro, nekuti metric uye mienzaniso inotsanangura "zvirinani." -Optimized zvirevo zvakachengeteka kuvimba. Vanogona kupfuudza diki dataset. - Optimization ndeyemahara. Inodhura chaiyo API mafoni, dzimwe nguva akawanda.
Chikwata chiri kuvaka tsigiro-tiketi classifier inonyora DSPy siginicha "tiketi -> chikamu" uye inopa mazana maviri akanyorwa matikiti uye metric yechokwadi. An optimizer anobva atora mashoma-mapfuti ekuratidzira anonakisa.
Kana kambani ikafamba kubva kune mumwe mupi weLLM ichienda kune mumwe, inodzokorodza DSPy optimizer yayo yemhando nyowani. Iyo hainyore neruoko-zvekukurudzira izvo zvakarongedzerwa kune yekare modhi.
Muongorori anomhanyisa OPRO-maitiro loop. Imwe modhi inokurudzira mazwi matsva ekuraira, mazwi ega ega anopihwa pabhenji yemasvomhu, uye iyo yakanyanya-zvibodzwa nhoroondo inodzokera muchikamu chinotevera chezvikumbiro.
Pombi yekudzoreredza-yakawedzerwa yemubvunzo-mhinduro inogadziridzwa kupera kusvika kumagumo, ine metric inoongorora mhinduro dzekupedzisira dzichipikisa mhinduro dzereferenzi. Zvese zviri zviviri nhanho yekunyora-mubvunzo uye nhanho yekupindura zvinovandudza pamwechete.
Kurumidza optimization iri kuita yakajairika chikamu cheLLM engineering, padivi pekuongorora masutu, pane kuda kuziva. Mibvunzo yakavhurika yasara. Mazano akakwenenzverwa anofambiswa sei pakati pemhando? Iwe unonyora sei mametrics ehunhu hunoenderana senge toni kana kubatsira? Iwe unodzivirira sei kuwandisa madiki dataset? Imwe tsvagiridzo inosanganisa kukurumidza optimization neyakajeka-tuning, uye DSPy inotsigira ese ari maviri. Mamodheru anochinja kazhinji, saka kugona kugadziridza otomatiki pakatarisana neyakagadzika bvunzo seti inogona kuve yakakosha kupfuura chero yakangwara yekumhanyisa. Kutsanangura mametrics akanaka uye mienzaniso inoramba iri basa revanhu.
Tsamba yemahara
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
Otomatiki nekukurumidza optimization inoshandisa algorithm kutsvaga zvirinani mirairo kana mashoma-epfuti mienzaniso. Inopa zvibodzwa zvemumiriri zvinopesana nemetric pane seti yemienzaniso pane kuvimba nemaoko-tweaked mazwi. Maitiro anozivikanwa anosanganisira APE, Google DeepMind's OPRO uye DSPy, yakavhurika-sosi chimiro kubva kuStanford. Pamwe chete vanoshandura huinjiniya nekukurumidza kubva pakufungidzira kuita inoyerwa, inodzokororwa maitiro.
Iyo optimizer inoda basa, mienzaniso uye metric saka inokwanisa kuhwina makwikwi emumiriri uye kuchengeta iwo akanakisa.
OPRO inoshandisa modhi yemutauro se optimizer. Inoona mirairo yapfuura nezvibodzwa zvavo uye inokurudzira mitsva.
Murairo wakaita zvakanaka kune imwe modhi pagiredhi-chikoro masvomhu. Mazwi anowanikwa nekutsvaga anogona kushamisa uye anogona kusaendeswa kune mamwe mamodheru.
Masiginicha anozivisa izvo module inofanirwa kuita maererano neinopinza uye inobuda minda. DSPy inobva yagadzirisa kukurumidza.
Matekisheni akabudirira anove mashoma-emienzaniso, kusanganisira yematanho emukati ausina kumbozvinyora iwe pachako.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Kurumidza Kushandura uye Regression Testing
Tekinoroji