Ngiyaxolisa, i-Prompt yam is not working!

Nangona iimpumelelo zakho ezilungileyo, i-LLM ayikho ngokufanelekileyo. Yintoni kufuneka utshintshe ngexesha elandelayo? Ngaba uqhagamshelane? Ukuguqulwa kwimodeli? Fine-tune? Yonke into ingaba iinketho ezifanelekileyo kwaye kukho inqaku yokufunda ezi zonyango.

Principle V: Follow Prompt-Fix Escalation Ladder

(Lezi zihlanganisa kwi-Principles of AI Engineering series: bheka iimpazamo1Ukucinga2Ukucinga3iimveliso4)

Xa isicelo esebenzayo asebenzayo njengoko kufuneka, ndiyabakhokela iifayile ezilandelayo ngexesha lokufuneka:

1. Expanding and rephrasing instructions.

2. Adding examples.

3. Adding an explanation field.

4. Using a different model.

5. Breaking up a single prompt into multiple prompts.

6. Fine-tuning the model.

7. Throwing the laptop out the window in frustration.

   
   

Kwiimeko ezininzi, umzekelo kwimibelelwano iya kuba ezahlukeneyo; Nangona kunjalo, ukufumana umzila we-default ukunciphisa ixesha kunye nokuvimbela umthamo wayo yokubhalisa. Le list ayenzelwe njenge-step rigid, kodwa njenge-guide rope eyenzelwe ukunceda.

Ngiya kuxhomekeke ngalinye iingxaki. Iintlobo ezintathu zokuqala zithunyelwa kwi-Prompt Engineering kwaye ziya kubhalwe ngakumbi kwi-chapter elandelayo. Iingxaki ze-Multi-prompt kunye ne-Fine-tuning ziya kubandakanyeka iingxaki ezizodwa.

Lightweight Approaches

Adding Instructions

Yinto yokuqala ukucwangcisa kubonisa kwakhona kwi-LLM njani nge-instructions prompt. Qinisekisa ukucwangcisa, ukucwangcisa, okanye ukucwangcisa imiyalelo.

Ngaba unemibuzo ukuguqula okanye ukuguqula iingcebiso ezininzi kwiindawo ezahlukeneyo ze-prompt - i-LLM ayidinga ukuguqula. Ngokuba izicwangciso ezininzi ezininzi ezininzi ezininzi, ukongeza ezininzi kwi-prompt okanye ekupheleni kwimiphumo ephakeme kakhulu (1Ukucinga2inguqulelo

Adding Examples

I-LLM zinxibelelana kakuhle kwi-in-context learning (i-input-output examples). Ziziphi na iimodeli ezincinane ezisetyenziswa; ezi zibonakalayo kakhulu "intelligent" ngoko kufuneka zihlanganisa kakhulu (3inguqulelo

Umzekelo wePrompt kunye ne-2-shot Inference (I-Language Detection):

Detect the language of the text and output it in the JSON format: {“language”: “name_of_language”}. If you don’t know the language, output “unknown” in the language field.

Example I:

Input: Hello

Output: {“language”: “English”}

Example II:

Input: EjnWcn

Output: {“language”: “Unknown”}

Text: {{text}}

Ngokuvamile uzisebenzisa 1-3 amaxabiso, nangona iimeko ezininzi ungayifaka. Kukho iziphumo ukuba imveliso iyahambisa kunye namanyathelo ezininzi (4), kodwa kunjalo i-cost ye-maintenance kunye ne-executive.

Adding an Explanation Field

I-LLM, njenge-umntu, ibonakalisa ukuba ibonakalise imiqondiso yayo. Ukongeza indawo ye- "i-explanation" kwi-JSON yakho ye-output kwaye i-output iya kuba kakuhle. Oku kuya kukunceda ukholelwa ukuba umzobo uyenza izixazululo ezithile kunye nokuguqula izicelo kunye neengxaki.

Ukusetyenziswa kweemveliso ye-inthanel ye-inthanel ye-inthanel ye-inthanel ye-inthanel ye-inthanel ye-inthanel ye-inthanel (5inguqulelo

Uyakwazi ukufumana i-AI-Chain-of-Thought I-Reasoning Prompt. Izixhobo apha kuya kuba ukusuka ngokufanelekileyo i-output. Unako kufuneka i-prompt yesibini okanye ikhowudi eyongezelelweyo ukuxhaswa imibuzo nge-CoT reasoning.

Changing the Model

Iimodeli ezahlukeneyo ziyafumaneka kwiintlobo ezahlukeneyo zophando. Iimodeli ye-OpenAI ye-o3 ziyafumaneka kwi-analyzing code, kodwa i-good old 4o inokukhawuleza ukuvelisa i-writing engcono nangona i-token engabizi. Indawo yobugcisa we-AI injineli ibandakanya izinzuzo kunye neengxaki zeemodeli ezikhoyo xa zithunyelwe kwaye zithunyelwe.

Ukusebenza ngempumelelo ngokukhawuleza nangokufanelekileyo xa unayo izifundo kunye ne-metric ezihambelana ne-"fitness" ye-modele ngamnye kumxholo.

Heavyweight Approaches

Yonke indlela ukuya kwakhona yaba ixabiso eliphantsi ngokufanelekileyo. Ngoku siphuma kwi-heavyweight fixes.

Breaking Up the Prompt

Ukuba umnqweno omnye akukwazi ukufumana umsebenzi - ngoko ke akufuneka inkqubo yeempawu ezimbini okanye ezininzi? Oku kungenziwa ukusebenza ngexesha elinye; iindlela ezimbini eziqhelekileyo zihlanganisa:

• Splitting the prompt by area of responsibility.

• Using a new prompt as a guardrail reviewing output of the previous one.

  
  

Zonke iinkqubo ziyafumaneka kwiIsahluko IIIkwaye iya kuxhomekeke ngexesha elandelayo ngakumbi.

Fine-Tuning

I-Fine-tuning yindlela engaphezulu kunyango kunyango ezininzi kunyango ezininzi. Kwiimeko ezininzi, ndisebenzisa njengendlela yokuqala.

Yintoni ndingathanda ukwamkela i-fine-tuning kwiimeko ezininzi? I-fine-tuning iyisiseko ye-machine learningisicelo Generative AINjengoko ke, kufuneka ukwamkela iimveliso ezininzi zebhanki kunye nenkqubo ye-ML ngaphandle kwe-AI generative kunye ne-AI generative kunye ne-overhead enkulu yokulawula iiprojekthi amancinci ukuya kwi-intermediate-size. Kwakhona inciphisa iinzuzo ze-AI eziphambili: ukucacisa kunye ne-logic ye-editing.

Qinisekisa i-fine-tuning xa:

• Zonke iinkqubo ezininzi ziye zikhusela ukufumana le mfuneko.
• Inkinga kubaluleke kakhulu kunye ne-specialized, kwaye iinkcukacha ze-LLM zangaphantsi.
• Ukuba unayo isicelo esikhulu kwaye ufuna ukunciphisa imali usebenzisa iimodeli ye-low-end.
• Ubuncinane ubuncinane kufuneka, ngoko iimpompo ezininzi ayidlulanga ngexesha.

Conclusion

Okwangoku, le nqaku ibonisa umzekelo weengxaki oya kufuneka uyenze xa iingxaki awukwazi ukusebenza ngokufanelekileyo. Okokuqala, uya kubandakanya indlela yokusebenza ngokufanelekileyo. Ukuba oku akufanelekileyo, nqakraza ukuguqulwa kwimodeli kwaye nqakraza ukuba oku kunyusa. Umgca elandelayo isebenzisa iingxaki ezininzi zokuxhumana. Ekugqibeleni, nqakraza ukuguqulwa okuphumelela ukuba zonke iindlela ezahlukileyo.

Ukuba ufumane le post - ukubhalisela ngaphezulu.