Qalisa futhi Uhlole Ukuhlolwa Kwenqubomgomo Yerobhothi Ehlanganiswe Nomzimba ye-GPT 6 Astra

Lokho okwenziwa yile phrojekthi

Le repository iqukethe ikhodi yokuhlola kanye nombiko wezobuchwepheshe onezilimi ezimbili mayelana nokusebenzisa i-GPT 6 Astra ekulawuleni amarobhothi. Iqhathanisa i-GPT 6 Astra Direct nenqubomgomo ehlanganisiwe ebuyekeza futhi, uma kudingeka, ilungise izenzo ze-π₀.₅ kuyo yonke imisebenzi eyishumi ye-RoboDojo, inamacala amahlanu ahambisanayo ngomsebenzi ngamunye.

Inqubomgomo ehlanganisiwe ibika i-Score emaphakathi engu-62.60 kanye nempumelelo engu-48%. Ukulungiswa kwe-GPT kwasetshenziswa ku-14.4% yezinyathelo zokulawula ezasetshenziswa. I-GPT 6 Astra Direct ithole impumelelo engu-26% kanye ne-Score emaphakathi engu-37.81 kuziqephu ezingu-48 ezazinamaphuzu endabuko atholakalayo. Womabili amazinga empumelelo asebenzisa wonke amacala akhethiwe angu-50. Iziqhathaniso zamamodeli asemthethweni ziyizinkomba zomphakathi ezilinganiswe kabusha, hhayi ukuphindaphindwa okusebenzisa ama-seed afanayo.

Lokhu kukhishwa kuyisithombe somthombo esingenawo umlando. Akuqukethe iziqinisekiso zokufinyelela, izichasiselo eziyimfihlo, izikhathi ezingaphekiwe zamamodeli, amalogi okuqalisa emisebenzi okuqala, amafa esifanisi, noma ama-checkpoint amamodeli. Isifanisi nezindawo zenqubomgomo ezidingekayo kufanele zitholakale ngokwehlukana kumaphrojekthi azo omthombo.

Qonda ukuhlelwa kwe-repository

• hybrid_rollout/robodojo/ iqukethe ukuhlanganiswa kwesifanisi, iseva yenqubomgomo, isilawuli, amakhono, kanye nomhleli wemisebenzi.
• hybrid_rollout/robolab/ iqukethe ukuhlanganiswa okuhlukile kwe-RoboLab.
• hybrid_rollout/report_site/ iqukethe umthombo wombiko onezilimi ezimbili, igalari, namathuluzi okupakisha.
• report_web/ iqukethe umbiko owakhiwe ngaphambilini kanye neziqeshana zawo zezihloko ezingu-21.
• public_results/ iqukethe imethadatha yamacala akhethiwe, amaphuzu, ama-seed, kanye nemvelaphi yedatha.
• licenses/ iqukethe izaziso zabanye abanikazi bamalayisensi.

Iwebhusayithi yokukhiqiza iqukethe igalari ephelele yamavidiyo futhi igcinwa kwenye i-repository. Iwebhusayithi ye-RoboDojo inikeza amavidiyo okuhlola angu-100 neziqeshana zegalari ezingu-43. Igalari ye-RoboLab inikeza okuqoshiwe okungu-150 okuhlanganisa izindlela ezintathu, imisebenzi eyishumi, kanye nezivivinyo ezinhlanu.

Buka kuqala umbiko onezilimi ezimbili

Indlela esheshayo yokuhlola lokhu kukhishwa ukuqala ukubuka kuqala kombiko wendawo usuka empandeni ye-repository. Iseva yokubuka kuqala isekela ama-byte range amavidiyo, ngakho umbiko ungadlala iziqeshana zawo ezifakiwe kahle.

python3 -m hybrid_rollout.report_site.preview --directory report_web --port 8768

Vula http://127.0.0.1:8768/ esipheqululini. IsiNgisi sikhethwa ngokuzenzakalela. Sebenzisa inkinobho yolimi noma wengeze ?lang=zh ukuze ushintshele esiShayina.

Ifolda emile ethi report_web/ nayo ingasingathwa ngeseva ejwayelekile ye-HTTP. I-GitHub Pages yomphakathi isebenzisa kuphela umbiko omile ohlanganisiwe nemidiya ekhethiwe; ayivezi i-API yezichasiselo noma izinsiza zokuqalisa imisebenzi.

Yakha kabusha i-frontend yombiko

Ukuze wakhe kabusha umbiko, sebenzisa i-Node.js ehambisana nenguqulo ye-Vite evaliwe. I-Node 24 yasetshenziswa kulokhu kukhishwa. Ngena kuhlelo lokusebenza lwe-frontend, faka okuncikile okuvalelwe, bese usebenzisa i-build yokukhiqiza.

cd hybrid_rollout/report_site/app
npm ci --ignore-scripts --no-audit --no-fund
npm run build

I-build iqinisekisa i-runtime eyabiwe egciniwe futhi idala i-dist/index.html ezimele ngokuphelele. Kopisha lelo fayela uye ku-report_web/index.html, ugcine i-data.json, scores.csv, citation.bib, amalayisensi, nemidiya eceleni kwalo.

Isithombe somthombo sokubhala somphakathi sifakiwe, ngakho isitolo sokuhlola esiyimfihlo asidingeki kulokhu kwakhiwa kabusha kwe-frontend. Ungasebenzisi i-build.py ngenhloso yokwakha kabusha i-frontend kuphela. Lelo thuluzi elingaxhunyiwe lokukhipha ubufakazi lidinga izingobo zomlando zendlela yokulandela zokuqala, ezingekho esidlangalaleni.

Lungiselela ukuhlola okusha

Ukuhlola okusha kudinga izindawo ezifakwe ngaphandle. Le repository ayisona isabalalisi sesifanisi esizimele.

1. Faka futhi ulungise i-Isaac Sim 5.1, umthombo namafa e-RoboDojo ahambisanayo, kanye ne-checkpoint ye-π₀.₅ OpenPI/JAX. I-RoboLab inezidingo zayo zendawo nokusetha. Thola konke lokhu kumaphrojekthi azo omthombo ngaphansi kwemigomo yawo.
2. Buyekeza hybrid_rollout/robodojo/SOURCE.json ukuze uhlonze izinguqulo zomthombo ezirekhodiwe kanye nobunikazi be-checkpoint.
3. Lungiselela indawo yamathuluzi ehlukile, indawo yesifanisi, indawo ye-OpenPI/JAX, i-checkpoint, kanye ne-Codex CLI. Gcina izindawo zesifanisi nezenqubomgomo zihlukene namathuluzi e-repository alula.
4. Kopisha noma ubuyekeze hybrid_rollout/robodojo/cluster.env.example. Setha izindlela zakho, isithombe se-container, ama-ID omsebenzisi naweqembu, izindlela zamafayela eziqinisekiso, kanye nobunikazi becala obucacile. Izindlela ezikokukhishwe komphakathi ziyizibonelo ezijwayelekile, aziwona ama-credential e-cluster asebenzisekayo.
5. Sebenzisa hybrid_rollout.robodojo.codex_backend.profiles ukuphatha ukuqinisekiswa kokusebenza okuhlukanisiwe. Hlola i-CLI usebenzisa --help ngaphambi kokungena noma ukwenza imisebenzi yokuqasha.
6. Sebenzisa public_results/evaluation_cases.json ukuvumelanisa umsebenzi, isigcawu, kanye ne-seed phakathi kwezinqubomgomo. Icala ngalinye liqukethe umphumela owodwa ohlanganisiwe kanye nomphumela owodwa we-Direct.

Imodeli imiswe ku-gpt-6-astra ngokucabanga kwe-xhigh. Udinga i-akhawunti yakho egunyaziwe noma ukufinyelela ku-gateway, futhi ukusebenzisa inqubomgomo noma imodeli kungase kudale izindleko.

Khetha indlela yokuhlola

Setha i-ROLLOUT_EVALUATION_METHOD ukuze ukhethe imodi yenqubomgomo:

ROLLOUT_EVALUATION_METHOD=pi05_plus_gpt

Sebenzisa i-pi05_plus_gpt kunqubomgomo ehlanganisiwe, evumela i-GPT ukuthi ibuyekeze futhi, uma kudingeka, ilungise izenzo ze-π₀.₅. Ukuze ukhethe inqubomgomo ye-Direct, sebenzisa:

ROLLOUT_EVALUATION_METHOD=gpt_only

gpt_only yisihlonzi sangaphakathi se-CLI se-GPT 6 Astra Direct. Ayethuli i-π₀.₅.

Indawo yokuqala yesiqukathi iku-hybrid_rollout/robodojo/cluster_entrypoint.sh. Isiqukathi ngasinye sisebenzisa icala elilodwa elicacile, silondoloze izinto esizikhiqizayo, bese sihlanza izingxenye zaso. Ifayela le-acp.py elingakhethwa liyiadaptha yokuthumela eqondene ne-SenseCore. Abasizi bokuthuthela imikhankaso abaqondene neqoqo badinga izinhlelo ezithile zokuphakela futhi abayona iresiphi yomphakathi yokuqalisa ngomyalo owodwa.

Qalisa izivivinyo ezingaxhunyiwe ku-inthanethi eziphindaphindekayo

Izivivinyo ezingaxhunyiwe ku-inthanethi zidinga okuncike kumathuluzi, kanti okuncike kusifanisi naku-OpenPI kufanele kufakwe ngokwehlukana ezindaweni zazo. Faka i-requirements-tools.txt endaweni ebonakalayo ezinikele, bese usebenzisa izivivinyo zombiko nezokuxilonga.

python3 -m unittest hybrid_rollout.report_site.test_i18n hybrid_rollout.report_site.test_editorial_revision -v
python3 -m pytest hybrid_rollout/robodojo/test_proposal_diagnostics.py hybrid_rollout/robodojo/test_prompt_context.py hybrid_rollout/robodojo/test_network_continue.py

Lezi zivivinyo azenzi ukufanisa noma izingcingo zemodeli. Izivivinyo ezincike kusitoreji sokuqala esabiwe, kusifanisi esiphathelene nobunikazi, noma ezintweni zomsebenzi zomlando azikwazi ukuphindaphindwa ngalesi sithombe somphakathi sodwa. Ukukhishwa akukaphindi kwenziwe izivivinyo ezikhokhelwayo zemodeli noma ze-GPU.

Imihlahlandlela yokuphepha nokuphakela

Ama-URL esango aku-codex_backend/profiles.py naku-codex_backend/config.toml asebenzisa izibambi-ndawo ezigodliwe ze-.invalid. Lungiselela ngokucacile indawo yakho egunyaziwe kuphrofayela ekhethiwe nakulungiselelo oluhambisanayo. Setha i-ROLLOUT_GATEWAY_NO_PROXY lapho inethiwekhi yakho ikudinga khona.

Ungalokothi ufake okhiye be-API noma i-auth.json ku-commit, futhi ungaguquli ikhaya lakho le-Codex elisebenzisanayo. Gcina imininingwane yokungena ngaphandle kwendawo yokugcina, inezimvume ezikhawulelwe. Izihlonzi zamaphrofayela zomlando zikhona ngenxa yokuhambisana kuphela; lokhu kukhishwa akulungiseleli inkampani noma insizakalo yokudlulisa.

Gcina amathuluzi ezichasiselo nawokubuyekeza kunethiwekhi yendawo ethembekile. Gcina ukubophela okuzenzakalelayo kwe-loopback, noma usebenzise ummeleli obuyisela emuva oqinisekisiwe ngaphambi kokwenza insizakalo ifinyeleleke ku-LAN. Gcina ukusetshenziswa kwamathuluzi e-Codex ne-RPC yesifanisi ngaphakathi kwendawo ehlukanisiwe nethembekile, futhi unganikeli ngezinhlelo ezingathembekile, idatha ye-RPC ehlelwe yaba uchungechunge, noma izingobo zomlando zemikhondo.

Ukusetshenziswa okuthuthukisiwe nezinkomba zocwaningo

Sebenzisa imethadatha yamacala eku-public_results/evaluation_cases.json ukuze ugcine ukuhambisana komsebenzi, kwesigcawu, kanye nembewu lapho uqhathanisa izinqubomgomo. I-manifest yegalari iqopha i-URL, usayizi, ne-SHA-256 yevidiyo ngayinye, okungasiza ukuqinisekisa imidiya elandiwe.

Ngombiko, igalari, neqoqo eliphelele lamavidiyo, vakashela umbiko negalari yamavidiyo. Ikhodi yomthombo yewebhusayithi yokukhiqiza iyatholakala endaweni yokugcina yewebhusayithi. Igalari ye-RoboLab inikeza okuqoshiwe okuhlukile kwe-RoboLab.

Ikhodi ephethwe yiphrojekthi inelayisensi ye-MIT. Izingxenye zezinkampani zangaphandle, amafonti, izimpahla zebenchmark, namalogo kugcina imigomo yako; bheka i-THIRD_PARTY_NOTICES.md. Isisindo samamodeli nezimpahla zesifanisi akusabalaliswa kabusha.

Isiphetho

Le ndawo yokugcina isetshenziswa kangcono njengokukhishwa komthombo wokuhlola nokubika: buka kuqala umbiko ofakiwe, yakha kabusha i-frontend yayo emile, sebenzisa izivivinyo ezitholakalayo ezingaxhunyiwe ku-inthanethi, futhi usebenzise imingcele yemvelo ebhaliwe ukulungiselela ukuqalisa okusha kwe-RoboDojo noma i-RoboLab. Ngaphambi kokuqala noma yimuphi umsebenzi wemodeli noma wesifanisi, thola okuncike kukho okudingekayo okuvela phezulu, lungiselela izinsizakalo ezigunyaziwe, qondanisa amacala nemethadatha yomphakathi, futhi uvikele imininingwane yokungena nezixhumi zokulawula.