Qeqesha Isihlukanisi Sombhalo Esisheshayo Nesibiza Kancane usebenzisa i-Tiny Classifiers ne-Ettin-17M

Okwenziwa Ama-Classifiers Amancane

tiny-classifiers iyindlela esebenzayo yokuqeqesha ama-classifier amancane ombhalo enzelwe imisebenzi ethile. Isebenzisa i-encoder ye-17-million-parameter Ettin-17M, ilungisa kahle zonke izisindo zayo, bese igcina umphumela njengemodeli ejwayelekile ye-Hugging Face.

Le phrojekthi yakhelwe imisebenzi yokuhlukanisa efana nokudlulisela imilayezo yasebhange ezindaweni ezifanele, ukuhlukanisa izindaba ngezigaba, ukuqaphela inhloso yomsebenzisi, ukuthola umuzwa, noma ukwabela imibhalo yezomthetho izigaba ezichazwe kusengaphambili. Uma ungenazo izibonelo ezanele ezilebulwe, imodeli yolimi enkulu ingasebenza njengothisha futhi ilebule idathasethi yokuqeqesha.

Ocwaningweni lwe-Banking77 olukurepository, i-Ettin-17M yafinyelela ukunemba okungu-91.5%, uma kuqhathaniswa no-92% we-Claude Opus. Ukuqeqesha kuthathe imizuzwana engu-42 ku-GPU eyodwa yokudlala, kanti ukusebenzisa imodeli kuthathe cishe ama-millisecond angu-5 ngomlayezo ngamunye. Ukuqeqesha kuphinde kwasebenza ku-CPU yelaptop cishe imizuzu engu-11 naku-GPU yocingo cishe imizuzu engu-21.

Ukunemba kwe-Banking77 uma kuqhathaniswa nesikhathi sokuqeqesha.

Izici Ezibalulekile

• Imodeli encane esebenza endaweni: I-Ettin-17M inamapharamitha ayizigidi ezingu-17 futhi ikhiqiza amaphuzu owodwa esigabeni ngasinye.
• Amalebula abantu noma othisha: Qeqesha ngamalebula aqinisekisiwe noma ucele i-Kimi K3 ilebule imilayezo engakalebulwa.
• Ukuqeqesha nge-GPU nange-CPU: Iskripthi sokuqeqesha sisekela zombili lezi zindawo.
• Umphumela wemodeli ejwayelekile: Ama-classifier aqeqeshiwe angalayishwa ngepayiphi ye-Hugging Face transformers.
• Ukulebula okuqhubekayo: Iskripthi sokulebula sisekela ukuzama futhi, ukwandisa noma ukunciphisa ukuxhumana ngokuya ngesimo, ukuphathwa kwezicelo ezihamba kancane, kanye nemikhawulo yabahlinzeki.
• Ukuhlola okwakhelwe ngaphakathi: Umbiko wokuqeqesha ubika ukunemba kokuhlolwa, kanti ukulebula kukathisha kungakala ukunemba kukathisha kudathasethi yokuhlola elebuliwe.
• Ukuhlaziywa kwezindleko: Iskripthi se-break-even silinganisela ukuthi ukulebula nokuqeqesha kuqala nini ukuba kushibhile kunokukhokhela ukuhlukaniswa nge-API.
• Ukulungiswa kwedatha okungasetshenziswa futhi: Le phrojekthi ingalungisa amadathasethi omphakathi ayishumi okuhlukanisa isebenzisa isakhiwo esisodwa esifanayo.

Indlela Esebenza Ngayo Iresiphi Yokuqeqesha

I-Ettin-17M iyi-encoder evulekile evela e-Johns Hopkins futhi ihlobene namamodeli ohlobo lwe-BERT. Ngokungafani nemodeli ekhiqizayo ekhiqiza umbhalo ithokheni ngethokheni, ifunda konke okufakwayo bese inikeza amaphuzu kuzo zonke izigaba ezitholakalayo.

I-repository yengeza okukhiphayo okukodwa esigabeni ngasinye futhi ilungisa kahle zonke izisindo zemodeli. Iresiphi yayo ezenzakalelayo isebenzisa usayizi weqoqo ongu-32, i-AdamW enezinga lokufunda elingu-1e-4, ukunciphisa isisindo okungu-0.01, ukufudumeza kokuqala okungu-5% okulandelwa ukwehla kwe-cosine, ukunqunywa kwe-gradient ku-1.0, kanye nokulahleka kwe-cross-entropy.

Emilayezweni ecishe ibe ngu-9,500, ukuqeqesha kudlula izikhathi eziyisithupha. Ngezibonelo ezingu-2,000 kuphela, sisebenzisa inani elifanayo lezinyathelo zokuthuthukisa, okulingana nokudlula okucishe kube ngu-28. I-README igcizelela ukuthi lezi zilungiselelo azilungiswanga ngokwehlukana emsebenzini ngamunye.

Ukufaka Nokulungisa

1. Faka i-uv

Izikripthi zisebenzisa i-uv ukudala indawo edingekayo nokufaka ngokuzenzakalelayo okuncikile ekuqaleni kokuqala. Faka i-uv usebenzisa imiyalelo yesistimu yakho yokusebenza, bese uvula i-terminal kumkhombandlela oyimpande we-repository.

2. Lungisa amadathasethi omphakathi

Qalisa iskripthi sokulungisa:

uv run recipe/prepare.py

Lokhu kulanda i-Banking77, i-CLINC150, i-MASSIVE ngesiNgisi nangesiFulentshi, i-TREC, i-AG News, i-SST-5, i-Financial PhraseBank, i-LEDGAR, kanye nedathasethi yezinkulumo zenzondo kuma-tweet evela ku-Hugging Face Hub.

Emsebenzini ngamunye, iskripthi sakha umkhombandlela ngaphansi kwe-data/tasks/. Sakha incazelo yomsebenzi, imilayezo yokuqeqesha engu-2,000, ingxenye enkulu yokuqeqesha equkethe izibonelo ezingafika kwezingu-150 esigabeni ngasinye, kanye nedathasethi yokuhlola.

3. Lungisa umhlinzeki wokulebula

Uma ufuna imodeli kathisha ikhiqize amalebula, setha ukhiye we-API womhlinzeki oyedwa noma ngaphezulu osekelwayo:

export OPENROUTER_API_KEY=...

Iresiphi iphinde ibone okhiye babahlinzeki abahlobene ne-Fireworks, Together, DeepInfra, kanye ne-Parasail. Hlola amanani amanje, izinqubomgomo zokusetshenziswa, nemigomo yemodeli yomhlinzeki ofanele ngaphambi kokuthumela idatha noma ukusebenzisa izimpendulo ezikhiqiziwe ekuqeqesheni.

Isifundo Esiyisisekelo: Qeqesha I-Classifier Ye-AG News

Isinyathelo 1: Khiqiza amalebula nge-Kimi K3

Ngemva kokulungisa amadathasethi, le­bula imilayezo yokuqeqesha engu-2,000 ye-AG News:

uv run recipe/label.py data/tasks/agnews

Ekusebenzeni okubikiwe, i-Kimi K3 ilebule izibonelo ezingu-2,000 cishe ngemizuzwana engu-20 ngezindleko ezicishe zibe ngu-$1. Isikhathi nezindleko zangempela zingahluka kuye ngomhlinzeki. Iskripthi singaqhubeka lapho besigcine khona, ngakho umsebenzi wokulebula ophazamisekile awudingi ukuqala ekuqaleni.

Isinyathelo 2: Lungisa kahle i-Ettin-17M ngamalebula kathisha

Dlulisela ifayela lamalebula akhiqiziwe kuskripthi sokuqeqesha bese ukhetha umkhombandlela wokugcina:

uv run recipe/train.py data/tasks/agnews --labels labels/agnews-train.jsonl --save models/agnews

Iskripthi silungisa kahle i-classifier, siyihlole kudathasethi yokuhlola, siphrinte ukunemba kokuhlola, bese sigcina imodeli ehambisana ne-Hugging Face ngaphansi kwe-models/agnews.

Isinyathelo 3: Qeqesha isisekelo samalebula abantu

Qalisa iskripthi esifanayo sokuqeqesha ngaphandle kwempikiswano yamalebula kathisha ukuze uqeqeshe ngamalebula abantu okuqala edathasethi:

uv run recipe/train.py data/tasks/agnews

Lokhu kuqhathanisa kusiza ukuveza ukuthi kungakanani ukusebenza okulahlekile ngesikhathi sokudluliswa kolwazi. Ekuqeqesheni kwesifundo se-repository, amalebula abantu akhiqize ukunemba okungu-89.2%, amalebula e-Kimi K3 akhiqize u-85.1%, kanti amalebula e-Qwen3.8-27B akhiqize u-81.8%.

Isinyathelo 4: Hlela umbhalo omusha

Layisha imodeli egciniwe nge-Hugging Face pipeline:

from transformers import pipeline

classify = pipeline(
    "text-classification",
    model="models/agnews",
    device="cpu"
)

result = classify(
    "NASA's new telescope finds water vapour on a distant planet"
)
print(result)

Umphumela uwuhlu oluqukethe ilebula ebikezelwe kanye nesikolo sayo. Kulesi sibonelo, isigaba esilindelekile yi-Sci/Tech. Imodeli ye-AG News okubikwe ngayo icubungule umlayezo cishe ngama-millisecond angu-6 ku-CPU.

Sebenzisa Ama-Classifiers Amancane Ngedatha Yakho

1. Dala ifolda yomsebenzi

Dala ifolda elandela ukwakheka okufanayo ne-data/tasks/agnews/. Umsebenzi ngamunye wangokwezifiso udinga amafayela e-task.json, train.jsonl, kanye ne-test.jsonl.

2. Chaza izigaba

Ifayela le-task.json kufanele libe negama, umusho owodwa womyalelo, kanye nohlu oluphelele lwamalebula:

{"name":"support-routing","instruction":"Assign each support message to the correct team.","labels":["billing","technical","account"]}

Sebenzisa amagama amalebula ahlukile futhi anencazelo. Umyalelo kufanele uchaze ngokucacile isinqumo sokuhlukanisa okufanele senziwe imodeli noma uthisha.

3. Faka amarekhodi okuqeqesha nawokuhlola

Gcina into eyodwa ye-JSON emgqeni ngamunye. Irekhodi elinelebula linesakhiwo esilandelayo:

{"id":"msg-001","text":"I was charged twice for my subscription.","label":"billing"}

Uma imodeli yothisha izolebula idatha yokuqeqesha, amarekhodi aku-train.jsonl angashiya ngaphandle inkambu ye-label:

{"id":"msg-002","text":"The application closes whenever I upload a file."}

Amarekhodi okuhlola kufanele abe namalebula aqinisekisiwe ukuze ukwazi ukukala ukunemba kukathisha nokomfundi. Iphrojekthi incoma ukuhlola mathupha imilayezo yokuhlola engamakhulu ambalwa ngaphambi kokwethemba imodeli ekukhiqizeni.

4. Lebula, qeqesha, bese uhlola

Faka ifolda yakho emiyalweni ejwayelekile:

uv run recipe/label.py data/tasks/support-routing
uv run recipe/train.py data/tasks/support-routing \
  --labels labels/support-routing-train.jsonl \
  --save models/support-routing

Ungaqeqesha ngokuqondile nangamalebula abantu ngokushiya ngaphandle i---labels. Ukuze usebenzise enye i-encoder esikhundleni se-Ettin-17M, dlulisa imodeli ngenketho ethi --model yeskripthi sokuqeqesha.

Linganisa Ukuthi Ukuqeqesha Kabusha Kuyazuzisa Nini

Qalisa ukuhlaziywa kwe-break-even ngemva kokulungisa imisebenzi:

uv run recipe/breakeven.py

Lesi skripthi silinganisela ukuthi mingaki imilayezo okufanele ihlukaniswe ngaphambi kokuba izindleko zokuqala zokulebula nokuqeqesha zibuyiswe, sibheka kokubili imali nesikhathi. Uma kuqhathaniswa ne-Je​​v zero-shot classification API, indawo yokufinyelela ku-break-even etholwe yirepository yayiphakathi kwemilayezo ecishe ibe ngu-50,000 no-500,000, kuye ngomsebenzi.

Uma umthamo wakho uncane noma ungaqinisekile, i-API ingaba ukukhetha okulula. Ukuqeqesha kabusha kuba okukhangayo kakhulu lapho umsebenzi uzinzile, idatha yokuhlola ikhona, futhi umthamo wokuhlukanisa uphezulu.

Amathiphu Athuthukile

Beka phambili ikhwalithi yamalebula

Ukuhlola kwathola ukuthi amalebula abaluleke kakhulu kunosayizi wemodeli. Ama-learning curves aqhubeke ethuthuka njengoba kwengezwa izibonelo eziningi ezinelebula. Umfundi oqeqeshwe uthisha ngokuvamile wahlala ngaphansi kukathisha wakhe ngamaphuzu ambalwa okunemba, kanti amalebula abantu ayengakwephula lokho kufinyelela okuphezulu.

Buyekeza izigaba ezingacacile, lungisa amaphutha ahlelekile okulebula, futhi wengeze izibonelo ezimele idatha ngaphambi kokushintsha ukwakheka noma ukulungisa ama-hyperparameter amaningi.

Hlola uthisha ngokwehlukana

Umfundi akakwazi ukulungisa ngokuthembekile uthisha osebenza kabi emsebenzini. Imiphumela yemisebenzi eyishumi yabonisa ukuthi amamodeli alebulwe uthisha alahlekelwa yinzuzo lapho uthisha uqobo ebuthaka. Gcina isethi yokuhlola ehlolwe abantu futhi ulinganise ukunemba kukathisha ngaphambi kokukhokhela ukulebula iqoqo elikhulu kakhulu.

Bheka ukwehluka kokuhlolwa

I-README ibika ukwehluka okucishe kube ngu-±0.5 wamaphuzu epesenti ngenxa yama-random seed, kanye no-±1.5 kuya ku-2 wamaphuzu ngenxa yosayizi wesethi yokuhlola. Amathebula ashicilelwe asebenzisa isilinganiso sez run ezintathu. Sebenzisa ama-run aphindaphindiwe namasethi okuhlola amakhulu ngokwanele lapho umehluko omncane wokunemba uthinta izinqumo zokuqalisa.

Hlonipha amalayisensi nemigomo yabahlinzeki

Qeqesha kuphela ngezimpendulo ezikhiqiziwe lapho ilayisensi yemodeli yothisha nemigomo yomhlinzeki ikuvumela lokho. I-repository yasebenzisa i-Kimi K3, kodwa amanye amamodeli nama-API ayakwenqabela ukusebenzisa imiphumela yawo ukuqeqesha enye imodeli. Amasethi edatha omphakathi nawo agcina amalayisensi awo; i-repository iyawalanda kunokuba iwaphinde iwasabalalise.

Hlola enye ihardware

Iskripthi esikhulu sokuqeqesha sisekela ukusebenza nge-CPU nange-GPU. I-repository iphinde ibe nokusetshenziswa okuhlukile kwefoni okuqeqesha nge-Chrome kusetshenziswa i-WebGPU ne-jax-js. Phatha ukuhamba komsebenzi kwefoni njengokusetshenziswa okukhethekile bese ulandela imiyalelo ekufolda yephrojekthi ethi phone/.

Sebenzisa isixazululo senkinga ebhaliwe ye-NixOS

Uma ukuqeqesha ku-NixOS kwehluleka ngoba /sbin/ldconfig ingatholakali, setha indlela yemitapo ye-CUDA ngaphambi kokuqalisa ukuqeqesha:

export TRITON_LIBCUDA_PATH=/run/opengl-driver/lib

Isiphetho

tiny-classifiers inikeza ukuhamba komsebenzi okufushane kokuguqula umbhalo onelebula ube yi-classifier esheshayo esebenza endaweni: lungisa umsebenzi, thola amalebula kubantu noma kuthisha, qeqesha kabusha i-Ettin-17M, hlola ngezibonelo eziqinisekisiwe, bese ulinganisela ukuthi ukucubungula endaweni kuyayidlula yini i-API ngokwezindleko.

Isifundo esibaluleke kakhulu wukutshala kumalebula athembekile nakudatha yokuhlola. Uma lokho sekukhona, i-encoder enamapharamitha ayizigidi ezingu-17 inganikeza ukunemba okuqinile kokuhlukanisa, ukucubungula okusezingeni lama-millisecond, kanye nezindleko zokusebenza zasendaweni ezibikezelwayo ngaphandle kokubiza imodeli enkulu kuwo wonke umlayezo.