Yenza ngokuzenzakalela Ukuqeqeshwa Kwangemva Kokuqeqeshwa kwe-LLM nge-AIBuildAI Agent

Uyini Umenzeli we-AIBuildAI LLM-Post-Train?

Umenzeli we-AIBuildAI LLM-Post-Train ungumenzeli ozithuthukisa ngokuphindaphindiwe, owenzelwe ukuqeqeshwa kwangemva kokuqeqeshwa okuzenzakalelayo nokwenziwe ngokwezifiso kwamamodeli amakhulu olimi. Esikhundleni sokusebenzisa indlela eyodwa yokuqeqesha engaguquki, uhlola umsebenzi, aklame ama-algorithm okungenzeka asetshenziswe ekuqeqeshweni kwangemva kokuqeqeshwa, aqoqe futhi ahlele idatha yokuqeqesha, abhale ikhodi, aqhube izivivinyo, ahlole imiphumela, bese ethuthukisa imodeli ngokuphindaphindiwe.

Le repository iqukethe ikhodi yomthombo yokusebenzisa, ukufunda, nokushintsha umenzeli. Yenzelwe izinhlangano nabacwaningi abadinga amasu okuqeqesha kwangemva kokuqeqeshwa ahambisana nedatha yobunikazi, imindeni yamamodeli, izinqubo zokuhlola, kanye nezinjongo zokusetshenziswa.

Ku-PostTrainBench, iphrojekthi ibika amaphuzu aphelele angu-46.6%, ibe sendaweni yokuqala phakathi kwamamodeli nabasebenzeli abahloliwe, futhi ilandelwa kuphela yesisekelo sochwepheshe ongumuntu esingu-51.1%.

Indlela Umenzeli Asebenza Ngayo

Ukusebenza kuqala ngomsebenzi kanye nenqubo yokufaka amaphuzu. Uhlelo lulungisa incazelo yomsebenzi engaguquki, lusungule isisekelo sokuqhathanisa, luqalise izivivinyo, luhlole umuntu ngamunye oqediwe lisebenzisa uhlelo olufanayo lokufaka amaphuzu, bese luguqula umphumela ongcono ube okungahanjiswa.

Amakhono amabili ahlukanisa lolu hlelo kumenzeli ojwayelekile wokubhala ikhodi:

• Uhlelo lolwazi: Abasebenzeli bangabuza Isizindalwazi Solwazi se-AIBuildAI nge-MCP. Imibhalo yamakhono ehleliwe ihlanganisa izinqubo zokuqeqesha kwangemva kokuqeqeshwa, izindlela, amadathasethi, kanye nezinhlaka.
• Ukusesha kwe-meta: Umenzeli we-meta angaphenya umsebenzi futhi abhale uhlelo lokusesha oluqondene nomsebenzi, esikhundleni sokusebenzisa ukwakheka kwezivivinyo okunqunywe kusengaphambili.

Ukusebenza ngakunye kuthola nendawo yokusebenza yewebhu ebukhoma, igrafu yokwenziwa ehlala ikhona, amayunithi omsebenzi abhalwe kwijenali, izinqubo ezingaphansi ezihlukaniswe nge-sandbox, kanye nezilawuli zezinsiza. Imicimbi yokusebenza igcinwa kuqoqo eliyimfihlo le-PostgreSQL ngaphansi kwe-~/.aibuildai/postgresql, okusekela ukusebenza okungaqhutshekwa futhi kudlalwe kabusha.

Izici Ezibalulekile

• Ukuklanywa nokuqhutshwa okuzenzakalelayo kwezivivinyo zokuqeqesha kwangemva kokuqeqeshwa kwama-LLM.
• Amasu okusesha aqondene nomsebenzi, okuhlanganisa ukusesha kwe-meta nezindlela ezimbalwa zokusesha ezimisiwe.
• Ukubuyiswa kolwazi lwezindlela, amadathasethi, izinhlaka, kanye nezinqumo zenqubo yomsebenzi.
• Abasebenzeli abanezinhlobo, izinhlelo ezinqunywa ngokucacile, kanye nokusesha okushumekiwe.
• Ukuhlolwa okuchazwa umsebenzi kanye nokubekwa kwabantu ngokwezinga ngendlela engaguquki.
• Ukuqapha okubukhoma okusekelwe esipheqululini ngesevisi yendawo yokusebenza yangakini.
• Ama-sandbox e-Bubblewrap kanye nemikhawulo ye-cgroup-v2 yememori nezinqubo.
• Imemori ekhiqizwa endaweni, ethathwe ekusebenzeni kwangaphambilini.
• Ukubhala okokuzikhethela kwamaphepha nokuhlanganiswa kwe-PDF ngemva kokuqoqwa.
• Ukusekelwa kwe-Claude namanye amamodeli atholakala ngezindawo zokugcina ezihambisana ne-Anthropic.

Izidingo Zohlelo

Ngaphambi kokufaka iphrojekthi, qinisekisa ukuthi umshini uhlangabezana nalezi zidingo:

• I-Linux ku-x86_64 eneseshini yomsebenzisi ye-systemd.
• Python 3.11 kuya ku-3.13.
• Amathuluzi okuhlanganisa e-C afaka i-gcc ne-make.
• Izihloko zokuthuthukisa ze-zlib ne-readline.
• bwrap, ehlinzekwa yiphakheji ye-bubblewrap, enezindawo zamagama zabasebenzisi abangavunyelwe amandla ezivuliwe.
• Ukufinyelela kunethiwekhi ekukhishweni kwe-GitHub naku-ftp.postgresql.org.
• Ukungena ku-Claude Code, ukhiye we-Anthropic API, noma enye imodeli etholakala ngendawo yokugcina ehambisana ne-Anthropic.
• I-GPU ye-CUDA iyanconywa, nakuba abasebenzeli bethola futhi basebenzise ihardware etholakalayo.

Ukusebenza kokuqala kwakha ukufakwa okuyimfihlo kwe-PostgreSQL kusuka ku-source tarball esemthethweni. Kuphinde kulande amakhophi anamavershini agxilile amathuluzi afana ne-ruff, tectonic, micromamba, caddy, kanye ne-PostgreSQL ku-~/.cache/aibuildai. Uhlelo lusebenzisa la makhophi alo esikhundleni sokufakwa okufanayo komshini.

Awudingi i-conda. Umenzeli udala izindawo ze-Program ngokufakwa kwe-micromamba okungokwakhe.

Hlola Ukunikezwa Kwe-cgroup-v2

Ukusebenza kuboshwe esihlahleni esisodwa se-cgroup-v2. Umphathi womsebenzisi we-systemd kufanele anikeze okungenani izilawuli ze-memory ne-pids. Hlola ukucushwa kwamanje ngale ndlela:

U=$(id -u)
cat /sys/fs/cgroup/user.slice/user-$U.slice/user@$U.service/cgroup.subtree_control

Okuphumayo kufanele kufake i-memory ne-pids. Uma kukhona okungekho, umlawuli kufanele engeze Delegate=pids memory cpu ku-drop-in engaphansi kwe-/etc/systemd/system/user@$U.service.d/, alayishe kabusha i-systemd, bese ekwenza ungene futhi.

Faka Umenzeli

1. Faka amaphakheji esistimu yokusebenza

Ku-Debian noma ku-Ubuntu, faka i-compiler, izihloko ezidingekayo, kanye ne-bubblewrap:

sudo apt-get install build-essential zlib1g-dev libreadline-dev bubblewrap

2. Clone i-repository

git clone https://github.com/aibuildai-inc/aibuildai-llm-posttrain-agent.git
cd aibuildai-llm-posttrain-agent

3. Dala indawo ebonakalayo ye-Python

python -m venv .venv
source .venv/bin/activate

4. Faka okuncikile okunamavershini agxilile

pip install -r requirements.lock
pip install -e . --no-deps

Ukufakwa okuhlelekayo kunikeza umyalo we-aibuildai. Sebenzisa ifayela lokukhiya ngoba i-claude-agent-sdk enenguqulo egxilile iyinguqulo eqinisekisiwe yale codebase.

Uma uhlela ukushintsha i-repository futhi usebenzise isango layo lekhwalithi, faka nokuncikile kokuthuthukisa:

pip install -r requirements-dev.lock

5. Lungisa ngokuzikhethela amathuluzi angawakho

Amathuluzi angalandwa futhi akhiwe ngesikhathi sokuqalisa kokuqala. Ukuze uwalungiselele kusenesikhathi, sebenzisa:

aibuildai setup

Lungiselela Ukufinyelela Kumodeli

Inketho 1: Sebenzisa ukungena ngemvume kwe-Claude Code

claude auth login

Inketho 2: Nikeza ukhiye we-API

export AIBUILDAI_API_KEY=sk-ant-...

Uma AIBUILDAI_API_KEY ingasethiwe, uhlelo luhlola ANTHROPIC_AUTH_TOKEN bese luhlola ANTHROPIC_API_KEY.

Sebenzisa enye indawo yokugcina ehambisana ne-Anthropic

Setha AIBUILDAI_BASE_URL ibe indawo yokugcina, bese ufaka ukhiye wayo ku-AIBUILDAI_API_KEY:

export AIBUILDAI_BASE_URL=https://openrouter.ai/api
export AIBUILDAI_API_KEY=your-key
export AIBUILDAI_SMALL_FAST_MODEL=provider/model-id

Izihlonzi zamamodeli ezingacacisiwe umhlinzeki eziqala ngo-deepseek- ziqondisa ngokuzenzakalelayo endaweni yokugcina ye-DeepSeek ehambisana ne-Anthropic. Izihlonzi zamamodeli e-OpenRouter zifaka isiqalo somhlinzeki futhi zidinga kokubili i-AIBUILDAI_BASE_URL ne-AIBUILDAI_SMALL_FAST_MODEL.

Uma indawo yokugcina isebenzisa amagama ahlukile ezigabeni zamamodeli e-Claude, uwamephu usebenzisa AIBUILDAI_HAIKU_MODEL, AIBUILDAI_SONNET_MODEL, kanye ne-AIBUILDAI_OPUS_MODEL. Sebenzisa AIBUILDAI_SUBAGENT_MODEL lapho ama-sub-agent kufanele asebenzise imodeli ehlukile kunemodeli emisiwe yendima yawo.

Izilungiselelo zesevisi ozikhethela zona

• AIBUILDAI_KB_BASE_URL ikhetha isevisi ye-knowledge-base MCP. Ngokuzenzakalelayo, iklayenti lisebenzisa i-snapshot yomphakathi engaguquki yephrojekthi.
• AIBUILDAI_KB_TOKEN inikeza i-bearer token uma isevisi yolwazi idinga ukuqinisekiswa.
• AIBUILDAI_WEB_PORT ikhetha i-port yesevisi yendawo yokusebenza yasendaweni.
• AIBUILDAI_WEB_NAME inikeza igama lesevisi yesibili yendawo yokusebenza kumsingathi ofanayo futhi idinga i-web port ecaciswe ngokusobala.

Dala Futhi Uqalise Ukuqhutshwa Okuyisisekelo

Ukuqhutshwa ngakunye kuchazwa ngefayela elilodwa le-YAML. Khiqiza ukucushwa kokuqala okunemibhalo usebenzisa:

aibuildai config > task.yaml

Ifayela elikhiqiziwe lifaka yonke inkambu yokucushwa, inani layo elizenzakalelayo, kanye nemibhalo yayo. Izigaba eziphezulu ezidingekayo yi-run, llm, work_units, search, kanye ne-resources; ifayela lokuqala lizigcwalisa.

Qalisa ukuqhutshwa usebenzisa:

aibuildai run task.yaml

Umyalo uphrinta i-URL yendawo yokusebenza efana ne-http://127.0.0.1:<port>/run/<run-id>. Yivule ukuze ulandele ukuqhutshwa bukhoma. Isiphequluli siyazikhethela, futhi ukuqhutshwa kuyaqhubeka ngokujwayelekile ngaphandle kwaso.

Qalisa Isibonelo se-GSM8K

Isibonelo esifakiwe senza i-post-training ye-Qwen/Qwen3-1.7B-Base ukuze ixazulule izinkinga zamagama zezibalo zasemabangeni aphansi ezivela ku-GSM8K. Ngokusho kwerepository, lesi sibonelo sithatha cishe amahora amabili ku-GPU eyodwa.

1. Yakha ifolda yomsebenzi

bash examples/gsm8k-qwen3-1.7b-base/setup.sh ~/gsm8k-example

I-setup ilanda cishe u-4 GB, okuhlanganisa izisindo zemodeli eyisisekelo kanye nama-split womabili e-GSM8K. Ibhalela isitatimende somsebenzi kanye ne-grader egxilisiwe ku-~/gsm8k-example/task/, futhi ifudumeza i-pip cache yamaphakheji afakwa ku-Program environment yokuqhutshwa.

Ifolda yomsebenzi ewumphumela iqukethe izisindo eziyisisekelo i-agent engazilungisa, izinkinga zokuqeqesha ezingu-7,473, i-grader, kanye ne-test split enezinkinga ezingu-1,319 efundwa yileyo grader.

2. Qala i-post-training

aibuildai run ~/gsm8k-example/task.yaml

Isibonelo sisebenzisa i-meta search, isikhathi sokuhlola esiyimizuzu engu-90, kanye namabhajethi ahambisanayo e-agent ngayinye. I-README incoma ukwandisa lawo mabhajethi ukuze kutholakale ukuqhutshwa okubalulekile.

3. Qonda imigomo yokuhlola

I-grader iyincazelo ephelele yempumelelo. Inikeza imodeli umbuzo olandelwa yisiqalo sempendulo, ikhiqize impendulo ngendlela greedy, bese ikhipha inombolo elandela i-#### yokugcina. I-README ibika ukuthi imodeli eyisisekelo ithola u-0.105 ezinkingeni zokuqala ezingu-200 zokuhlola ngoba imodeli eyisisekelo engakaqeqeshwa ekuqaleni ayilandeli ifomethi yempendulo edingekayo.

4. Hlola imodeli yokugcina

Okulethwayo yifolda yemodeli ehlanganisiwe. Faka igama lefolda yokuqhutshwa ekhiqiziwe esikhundleni se-<run-home>, bese uhlola zonke izibonelo zokuhlola usebenzisa i-Program environment yokuqhutshwa:

H=~/gsm8k-example/playground/gsm8k-qwen3-1.7b-base/<run-home>
$H/program-environment/env/bin/python ~/gsm8k-example/task/grade/evaluate.py --model-path $H/deliverable/final_model --limit -1

Sebenzisa I-Agent Ngomsebenzi Wakho

1. Kopisha examples/gsm8k-qwen3-1.7b-base.yaml, noma ukhiqize ifayela lokuqala eliphelele usebenzisa aibuildai config.
2. Dala ifolda yomsebenzi equkethe konke okungasetshenziswa ukuqhutshwa.
3. Faka incazelo yomsebenzi efundekayo, njenge-README, isitatimende somsebenzi, noma iphepha, kanye nawo wonke amafayela edatha adingekayo.
4. Setha run.data_root kuleyo folda.
5. Buyekeza izilungiselelo zemodeli, zokusesha, zebhajethi, zamayunithi omsebenzi, nezinsiza.
6. Qalisa ifayela le-YAML usebenzisa aibuildai run.

Ngesikhathi se-setup, uhlelo lufunda yonke ifolda yomsebenzi, lugcine isitatimende singaguquki njenge-README yokuqhutshwa, bese lwakha uhlelo lwamaphuzu lokuqhutshwa kuso. Ngakho-ke ifolda yomsebenzi kufanele iphelele futhi ichaze ngokusobala lokho okulindelekile okulethwayo kanye nendlela yokuhlola.

Ukucushwa kusekela futhi run.setup_from_scratch kanye ne-run.task_prompt. Lezi zilungiselelo zivumela i-Setup ukuthi ithole idatha nokuhlola esitatimendeni somsebenzi esikhundleni sokuqala ngefolda yomsebenzi elungiselelwe.

Izinketho Ezibalulekile Zokucushwa

Ukukhetha imodeli

llm.default.model ikhetha imodeli eyisisekelo esetshenziswa ama-agent. Ukucabanga kungalungiswa ngezilungiselelo ezifana ne-effort kanye ne-max_thinking_tokens. Sebenzisa i-llm.by_role ukunamathisela amamodeli ezindimeni ngazinye, noma i-llm.auto ukuze i-router ikhethe ngokuya ngendima.

Indlela yokusesha

search.kind ilawula isu lokuqhuba i-experiment. Amanani asekelwayo achazwe ku-README afaka:

• meta, lapho i-meta agent ibhala khona uhlelo lokusesha oluqondene nomsebenzi.
• nb_tree, usesho lwesihlahla lwe-experiment yonke.
• nb.
• tree, okuyinto ezenzakalelayo ku-general starter.
• linear.
• parallel.

Okuqukethwe kwe-search.input kuncike esu elikhethiwe. I-meta search yamukela inketho ye-report, kanti amasu angaguquki asebenzisa amapharamitha awo futhi angafaka i-parallel ne-early_stopping.

Imikhawulo yesikhathi neyokusetshenziswa kwemali

run.budget.wall_clock_minutes okuzenzakalelayo imizuzu engu-1,440. Uma umkhawulo usufinyelelwe, akukho kuhlola okusha okuqalayo, kodwa umsebenzi oqhubekayo nokulethwa kuvunyelwa kuqede. run.budget.cost_usd inikeza umkhawulo ongaphoqeleli wokusetshenziswa kwemali ezingeni le-run.

Into ngayinye engaphansi kwe-work_units inesabelomali sayo. I-work unit ingasebenzisa imizuzu engaguquki, isikhathi esikalwe ngokujula, noma ubude besikhathi obulindelekile obune-slack.

Imikhawulo yezinsiza

Sebenzisa resources.work_unit.cpu_max_cores kanye ne-resources.work_unit.memory_max_gb ukubeka imikhawulo eqinile ye-CPU nememori ku-work unit ngayinye. Iyunithi ngayinye isebenza ku-sandbox ye-bubblewrap enemikhawulo ye-cgroup kumemori nasezinqubweni, kanti ifolda yomsebenzi ifakwa ngemodi yokufunda kuphela.

Amathiphu Athuthukile

Vumela ukubuyiswa kolwazi

Faka ukumiswa kwe-kb MCP lapho ufuna ama-agent axhumane ne-AIBuildAI Knowledge Base. Iklayenti linikeza amathuluzi okusesha ama-skill, ukulayisha i-skill, nokufunda izinkomba. I-corpus ihlanganisa ukuqondiswa kokuhamba komsebenzi, amakhadi endlela yokwenza, amakhadi edathasethi, namakhadi e-framework.

Isevisi yomphakathi ezenzakalelayo iyi-snapshot ye-corpus engaguquki. Ukuze usingathe eyakho inguqulo, faka i-open-source knowledge-base repository bese ukhomba AIBUILDAI_KB_BASE_URL kuyo.

Khetha i-meta search nge-topology eqondene nomsebenzi

I-fixed tree search ayikwazi ukuveza ngokwemvelo yonke imisebenzi yokuhamba komsebenzi, njengokuqhuba i-sweep ngaphambi kokuqeqesha, ukubamba umqhudelwano ngemva kwalokho, noma ukuhlanganisa amamodeli amaningi. Nge-search.kind: meta, i-meta agent iphenya umsebenzi, ingaqhuba ama-pilot, ibheke ama-design skill afakiwe kanye ne-knowledge base, bese ibhala i-Python search package.

Iphakheji ekhiqiziwe yakhiwe ngezinto ezintathu eziyisisekelo:

• Agent: Iseshini ye-LLM enamathuluzi, izinqubomgomo, kanye nokuphumayo okunohlobo olucacisiwe.
• Program: Ukubala okunqunywa kusengaphambili ku-worker ephethwe enemikhawulo yezinsiza ezinikezelwe.
• Search: I-orchestrator ekwazi ukudala nokuxhumanisa ama-Agent, ama-Program, nama-Search angaphakathi isebenzisa okokufaka nokuphumayo okunohlobo olucacisiwe.

Iphakheji iyaqinisekiswa futhi ingabuyekezwa ngokuzikhethela ngombiko womklamo we-LaTeX. Ngemva kwalokho isebenza njenge-child search ku-run efanayo enomlando ogciniwe. Kungakhathaliseki i-topology, inani le-run lihlala liyisikolo esihle kakhulu esitholwe yinoma iyiphi i-node.

Yakha imemori yendawo esebenza kabusha

Ngemva kokuqongelela ama-run omsebenzi, hlanganisa izifundo zawo zibe idokhumenti yememori yendawo ehlelekayo:

aibuildai memorize task.yaml

Setha memory.enable ukuze ufake leyo memori kuwo wonke ama-agent. Ulwazi olufundwe kuma-run akho luhlala emshinini wakho futhi alufakwa ku-knowledge base ekude.

Lawula amathuluzi angaphandle

Sebenzisa disallowed_tools ukuvimbela yonke indima ekusebenziseni amathuluzi akhelwe ngaphakathi akhethiwe. Isibonelo, umsebenzi ungavimbela WebSearch ne-WebFetch lapho imithetho yokuhlola idinga ukuhamba komsebenzi okungaxhunyiwe ku-inthanethi.

Engeza ukuhlanganiswa kwe-MCP

Ukumiswa kwe-mcps kungavumela i-knowledge base, ukuthunyelwa komncintiswano we-Kaggle, noma amaseva e-MCP angokwezifiso. Ukuhlanganiswa kwe-Kaggle kusekela isilungiselelo sokuthunyelwa okuphezulu ngaphansi kwe-submission.max_submissions.

Khiqiza iphepha

Vumela writer.enable ukuze kubhalwe uhlaka lwephepha ngemva kokuhlanganisa, noma sebenzisa umyalo wephepha kufolda ye-run eqediwe:

aibuildai write-paper <run-dir>

Umyalo ubhala uhlaka lwephepha kusukela ku-run eqediwe bese uluhlanganisa lube yi-PDF.

Qinisekisa izinguquko zomthombo

Lolu chungechunge lwephrojekthi alunayo i-test suite. Isango le-repository yalo ngu:

./check.sh

Isango lisebenzisa i-pyright, i-ruff, i-semgrep, kanye nokuhlola umehluko we-lock-file. Faka requirements-dev.lock ngaphambi kokulisebenzisa.

Qonda Okuphumayo Kwe-Run

I-run ngayinye idala ifolda yasekhaya ngaphansi kwe-playground root emisiwe:

{playground_root}/<task-name>/<timestamp>_<run-id>/
  run_config.json
  public/
  private/
  deliverable/
  workspace/
  program-environment/
  resource_samples.jsonl

• run_config.json iqukethe ukumiswa kokugcina kanye ne-run ID esetshenziselwa ukuqhubeka.
• public/ iqukethe i-README egcinwe ingaguquki kanye nedatha elungiselelwe ama-candidate yi-Setup.
• private/ iqukethe i-score program, i-metric manifest, umzamo oyisisekelo, kanye nerekhodi lomnqobi.
• deliverable/ iqukethe iphakheji yemodeli esebenzisekayo noma omunye umphumela ochazwe umsebenzi.
• workspace/ iqukethe imizamo, ama-artifact, idatha yesikhashana, kanye namafolda omsebenzi e-search ngayinye.
• program-environment/ iqukethe i-Program environment ehlukanisiwe ye-run.
• resource_samples.jsonl irekhoda izilinganiso zezinsiza ze-host eziqoqwe ngesikhathi sokwenza.

Isiphetho

I-AIBuildAI LLM-Post-Train Agent ihlinzeka ngohlelo oluphelele lokuklama, ukusebenzisa, ukukala, nokuthuthukisa izinqubo zokuqeqeshwa kwangemva kokuqeqeshwa kwemodeli. Qala ngesibonelo esinikeziwe se-GSM8K ukuze uqinisekise ukufakwa futhi ufunde indlela yokusebenzisa uhlelo. Bese uchaza ifolda yomsebenzi ecacile nomhloli, ukhethe isu lokusesha elifanele, usethe amabhajethi anengqondo nemikhawulo yezinsiza, bese usebenzisa indawo yokusebenza ebukhoma ukulandelela izivivinyo eziba khona.

Ukuze uthole imininingwane ejulile mayelana nesakhiwo sohlelo, bheka imibhalo yesakhiwo sohlelo ekhoselweni. Iphrojekthi isatshalaliswa ngaphansi kwe-Apache License 2.0.