Ukwakha i-AI ekhethekile ngaphandle kokudela ubuhlakani: I-Nova Forge data mixing in action

nimda March 2, 2026

0 8 7 minutes read

Ukwakha i-AI ekhethekile ngaphandle kokudela ubuhlakani: I-Nova Forge data mixing in action

Amamodeli ezilimi ezinkulu (ama-LLM) enza kahle emisebenzini evamile kodwa alwela umsebenzi okhethekile odinga ukuqonda idatha yobunikazi, izinqubo zangaphakathi, namagama aqondene nemboni ethile. I-Supervised fine-tuning (SFT) ivumelanisa ama-LLM nalezi zimo zenhlangano. I-SFT ingasetshenziswa ngezindlela ezimbili ezihlukene: I-Parameter-Efficient Fine-Tuning (PEFT), ebuyekeza kuphela isethi encane yamapharamitha angamamodeli, inikeza ukuqeqeshwa okusheshayo nezindleko eziphansi zokubala kuyilapho kugcinwa ukuthuthukiswa kokusebenza okunengqondo; I-SFT yezinga eligcwele, ebuyekeza wonke amapharamitha angamamodeli kunesethi engaphansi futhi ehlanganisa ulwazi lwesizinda olwengeziwe kune-PEFT.

I-SFT yezinga eligcwele ivamise ukubhekana nenselele: ukukhohlwa okuyinhlekelele. Njengoba amamodeli efunda amaphethini aqondene nesizinda esithile, alahlekelwa amakhono ajwayelekile ahlanganisa ukulandela imiyalelo, ukucabanga, nolwazi olubanzi. Izinhlangano kufanele zikhethe phakathi kochwepheshe besizinda kanye nobuhlakani obujwayelekile, obukhawulela insiza yamamodeli kuzo zonke izimo zokusetshenziswa kwebhizinisi.

I-Amazon Nova Forge ibhekana nenkinga. I-Nova Forge iyisevisi entsha ongayisebenzisela ukuzakhela amamodeli emingcele yakho usebenzisa i-Nova. Amakhasimende e-Nova Forge angaqala ukuthuthukiswa kwawo kusukela ezindaweni zokuhlola amamodeli angaphambi kwesikhathi, ahlanganise idatha yokuphathelene nedatha yokuqeqeshwa ekhethiwe ye-Amazon Nova, futhi asingathe amamodeli awo angokwezifiso ngokuphephile ku-AWS.

Kulokhu okuthunyelwe, sabelana ngemiphumela evela ekuhloleni okuphelele kwethimba le-AWS China Applied Science kwe-Nova Forge lisebenzisa umsebenzi wokuhlukanisa oyinselele we-Voice of Customer (VOC), omakiwe ngokumelene namamodeli omthombo ovulekile. Ngokusebenzisana namasampula okuphawula kwamakhasimende angaphezu kuka-16,000 kuwo wonke amalebula ayinkimbinkimbi anamazinga amane aqukethe izigaba zamaqabunga angu-1,420, sibonisa indlela indlela ye-Nova Forge yokuxuba idatha enikeza ngayo izinzuzo ezimbili:

Izinzuzo zokusebenza komsebenzi wangaphakathi kwesizinda: ukuzuza ukuthuthukiswa kwamaphuzu angu-17% F1
Amakhono ajwayelekile agciniwe: ukugcina amaphuzu aseduze nesisekelo se-MMLU (Ukuqonda Ulimi Oluningi Lwemisebenzi Eningi) kanye namakhono alandela imiyalelo ngemva kokuhluzwa

Inselele: ukuhlelwa kwempendulo yekhasimende lomhlaba wangempela

Cabangela isimo esijwayelekile enkampanini enkulu ye-ecommerce. Ithimba lomuzwa wamakhasimende lithola izinkulungwane zamazwana amakhasimende nsuku zonke anempendulo enemininingwane ehlanganisa ikhwalithi yomkhiqizo, okuhlangenwe nakho kokulethwa, izinkinga zokukhokha, ukusebenziseka kwewebhusayithi, kanye nokusebenzisana kwesevisi yamakhasimende. Ukuze basebenze kahle, badinga i-LLM engahlukanisa ngokuzenzakalelayo amazwana ngamunye abe izigaba ezingenziwa ngokunemba okuphezulu. Ukuhlelwa ngakunye kufanele kucace ngokwanele ukuze kudluliselwe udaba eqenjini elifanele: ezokuthutha, ezezimali, ukuthuthukiswa, noma isevisi yamakhasimende, futhi kucuphe ukuhamba komsebenzi okufanele. Lokhu kudinga ubuchwepheshe besizinda.

Nokho, yona le LLM ayisebenzi yodwa. Kuyo yonke inhlangano yakho, amaqembu adinga imodeli ukuze:

Khiqiza izimpendulo ezibheke amakhasimende ezidinga amakhono okuxhumana ajwayelekile
Yenza ukuhlaziya idatha edinga ukucabanga kwezibalo nokunengqondo
Amadokhumenti asalungiswa kulandela imihlahlandlela ethile yokufometha

Lokhu kudinga amakhono ajwayelekile jikelele– ukulandela imiyalelo, ukucabanga, ulwazi kuzo zonke izizinda, nokukhuluma ngokushelelayo.

Indlela yokuhlola

Uhlolojikelele lokuhlola

Ukuhlola ukuthi i-Nova Forge ingakwazi yini ukuletha kokubili okukhethekile kwesizinda namandla avamile, siklame uhlaka lokuhlola okukabili olukala ukusebenza kuzo zonke izinhlangothi ezimbili.

Ngokusebenza okuqondene nesizinda, sisebenzisa i-a I-real-world Voice of Customer (VOC) idathasethi kususelwa kuzibuyekezo zangempela zekhasimende. Idathasethi iqukethe amasampula okuqeqeshwa ayi-14,511 kanye namasampula okuhlola angama-861, abonisa idatha yezinga lokukhiqiza lebhizinisi. Idathasethi isebenzisa i-taxonomy yamazinga amane lapho iLeveli 4 imele izigaba zamaqabunga (okuhloswe ngakho kokugcina). Isigaba ngasinye sihlanganisa incazelo echazayo yobubanzi baso. Izigaba eziyisibonelo:

Ileveli 1	Izinga 2	Ileveli 3	Ileveli 4 (isigaba samaqabunga)
Ukufakwa – ukucushwa kohlelo lokusebenza	Umhlahlandlela wokusetha wokuqala	Inqubo yokusetha	Easy isethaphu umuzwa: Izici zenqubo yokufaka kanye nezinga eliyinkimbinkimbi
Ukusetshenziswa – isipiliyoni sehadiwe	Ukusebenza kombono wasebusuku	Ikhwalithi yesithombe yokukhanya okuphansi	Ukucaca kombono wasebusuku: Imodi yombono wasebusuku ikhiqiza izithombe ekukhanyeni okuphansi noma ezimeni ezimnyama
Ukusetshenziswa – isipiliyoni sehadiwe	Umsebenzi we-Pan-tilt-zoom	Ikhono lokuzungezisa	360-degree ukuzungezisa: Ikhamera ingazungezisa amadigri angu-360 agcwele, inikeze ukumbozwa okuphelele kwe-panoramic
Inqubomgomo kanye nezindleko zangemuva kokuthengisa	Inqubomgomo yokubuyisela nokushintshanisa	Buyisa ukusebenza kwenqubo	Ukubuyisela umkhiqizo kuqediwe: Ikhasimende eliqalwe futhi laqeda ukubuyisela umkhiqizo ngenxa yezinkinga zokusebenza

Idathasethi ibonisa ukungalingani kwekilasi okwedlulele okufana nezimo zempendulo zamakhasimende zomhlaba wangempela. Isithombe esilandelayo sibonisa ukusatshalaliswa kwekilasi:

Ngenxa yalokho, idathasethi ibeka inselele enkulu ekunembeni kwezigaba.

Ukuze sihlole amakhono enhloso evamile, sisebenzisa umphakathi ukuhlukaniswa kwesethi yokuhlola kwe I-MMLU (Ukuqonda Okukhulu Kolimi Lwemisebenzi Eningi) (wonke amasethi angaphansi). Ukuhlolwa kuhlanganisa izifundo kubuntu, isayensi yezenhlalo, isayensi eqinile, nezinye izindawo ezibalulekile ukuthi abanye abantu bazifunde. Kulokhu okuthunyelwe, i-MMLU isebenza njengommeleli we ukugcinwa kwamandla jikelele. Siyisebenzisela ukukala ukuthi ukulungisa kahle okugadiwe kuyayithuthukisa yini ukusebenza kwesizinda ngezindleko zokuziphatha kwemodeli eyisisekelo eyehlisa isithunzi, kanye nokuhlola ukusebenza kahle kokuxutshwa kwedatha ye-Nova ekunciphiseni ukukhohlwa okuyinhlekelele.

Into	Incazelo
Isamba samasampuli	15,372 izibuyekezo zamakhasimende
Ukulandelana kwelebula	Ukuhlukaniswa kwamazinga ama-4, izigaba eziyi-1,420 sezizonke
Isethi yokuqeqesha	14,511 amasampula
Isethi yokuhlola	861 amasampula
I-MMLU Benchmark all (ukuhlukaniswa kokuhlola)	14,000 amasampula

Ukuhlolwa komsebenzi ongaphakathi kwesizinda: izwi lokuhlukaniswa kwekhasimende

Ukuze siqonde ukuthi i-Nova Forge isebenza kanjani ezimeni zebhizinisi langempela, siqale sihlole ukunemba kwemodeli kumsebenzi wokuhlukanisa we-VOC ngaphambi nangemuva kokushuna kahle okugadiwe. Ngale ndlela, singakwazi ukulinganisa izinzuzo zokujwayela isizinda ngenkathi sisungula isisekelo sokuhlaziywa kokuqina okwalandela.

Ukuhlola imodeli eyisisekelo

Siqala ngo-a ukuhlolwa kwemodeli eyisisekelo ukuhlola ukusebenza kwangaphandle kwebhokisi kumsebenzi wokuhlukanisa we-VOC ngaphandle kwanoma yikuphi ukulungisa kahle okuqondene nomsebenzi othile. Lokhu kusetha kusungula ikhono lemodeli ngayinye lokusingatha ukuhlukaniswa kwe-granular kakhulu ngaphansi kwemikhawulo eqinile yefomethi yokuphumayo. Lokhu okulandelayo kusetshenziselwa umsebenzi wokuhlukanisa we-VOC:

# Role Definition

You are a rigorous customer experience classification system. Your sole responsibility is to map user feedback to the existing label taxonomy at Level 1 through Level 4 (L1–L4). You must strictly follow the predefined taxonomy structure and must not create, modify, or infer any new labels.

## Operating Principles

### 1. Strict taxonomy alignment

All classifications must be fully grounded in the provided label taxonomy and strictly adhere to its hierarchical structure.

### 2. Feedback decomposition using MECE principles

A single piece of user feedback may contain one or multiple issues. You must carefully analyze all issues described and decompose the feedback into multiple non-overlapping segments, following the MECE (Mutually Exclusive, Collectively Exhaustive) principle:

- **Semantic singularity**: Each segment describes only one issue, function, service, or touchpoint (for example, pricing, performance, or UI).

- **Independence**: Segments must not overlap in meaning.

- **Complete coverage**: All information in the original feedback must be preserved without omission.

### 3. No taxonomy expansion

You must not invent, infer, or modify any labels or taxonomy levels.

## Label Taxonomy

The following section provides the label taxonomy: {tag category}. Use this taxonomy to perform L1–L4 classification for the original VOC feedback. No taxonomy expansion is allowed.

## Task Instructions

You will be given a piece of user feedback: {user comment}. Users may come from different regions and use different languages. You must accurately understand the user's language and intent before assigning labels.

Refer to the provided examples for the expected labeling format.

## Output Format

Return the classification results in JSON format only. For each feedback segment, output the original text along with the corresponding L1–L4 labels and sentiment. Do not generate or rewrite content.

```json

[

{

"content": "",

"L1": "",

"L2": "",

"L3": "",

"L4": "",

"emotion": ""

}

]

```

Ukuhlola imodeli eyisisekelo, sikhethe:

Imodeli	Ukunemba	Khumbula	I-F1-Score
I-Nova 2 Lite	0.4596	0.3627	0.387
Qwen3-30B-A3B	0.4567	0.3864	0.394

Izikolo ze-F1 ziveza lokho I-Nova 2 Lite ne-Qwen3-30B-A3B zibonisa ukusebenza okufanayo kulo msebenzi oqondene nesizinda, womabili amamodeli athola amaphuzu angu-F1 eduze kuka-0.39. Le miphumela iphinda igqamise ubunzima obukhona bomsebenzi: ngisho namamodeli esisekelo aqinile alwa nokuhlukaniswa kwelebula ecolisekile uma kungekho datha eqondene nesizinda enikeziwe.

Ukushuna kahle okugadiwe

Sibe sesifaka isicelo i-full-parameter supervised fine-tuning (SFT) usebenzisa idatha ye-VOC yekhasimende. Wonke amamodeli acushwe kahle kusetshenziswa idathasethi efanayo kanye nokulungiselelwa kokuqeqeshwa okufanayo ukuze kuqhathaniswe okulungile.

Ingqalasizinda yokuqeqesha:

Ukuqhathanisa kokusebenza komsebenzi wesizinda

Imodeli	Idatha Yokuqeqesha	Ukunemba	Khumbula	I-F1-Score
I-Nova 2 Lite	Lutho (isisekelo)	0.4596	0.3627	0.387
I-Nova 2 Lite	Idatha yekhasimende kuphela	0.6048	0.5266	0.5537
Qwen3-30B	Idatha yekhasimende kuphela	0.5933	0.5333	0.5552

Ngemva kokulungisa kahle idatha yekhasimende kuphela, I-Nova 2 Lite izuza ukuthuthukiswa okukhulu kokusebenzane-F1 ekhuphuka isuka ku-0.387 iye ku-0.5537—inzuzo ephelele yamaphuzu angu-17. Lo mphumela ubeka imodeli ye-Nova esigabeni esiphezulu salo msebenzi futhi wenza ukusebenza kwayo kuqhathanise nokwemodeli yomthombo ovulekile ye-Qwen3-30B ecushwe kahle. Le miphumela iqinisekisa ukusebenza kwe Nova ipharamitha egcwele ye-SFT imisebenzi enzima yokuhlukanisa amabhizinisi.

Ukuhlolwa kwamakhono ajwayelekile: ibhentshimakhi ye-MMLU

Amamodeli ashunwe kahle wesigaba se-VOC avame ukuthunyelwa ngalé komsebenzi owodwa futhi ahlanganiswe nokugeleza komsebenzi okubanzi kwebhizinisi. Ukugcina amakhono enhloso ejwayelekile kubalulekile. Amabhentshimakhi ajwayelekile embonini afana ne-MMLU ahlinzeka ngendlela esebenzayo yokuhlola amakhono enhloso evamile kanye nokuthola ukukhohlwa okuyinhlekelele kumamodeli ashunwe kahle.

Kumodeli ye-Nova ecushwe kahle, i-Amazon SageMaker HyperPod inikeza izindlela zokupheka eziphuma ngaphandle kwebhokisi ezilula ukuhlolwa kwe-MMLU ngokucushwa okuncane.

Imodeli	Idatha yokuqeqesha	I-VOC F1-Score	Ukunemba kwe-MMLU
I-Nova 2 Lite	Lutho (isisekelo)	0.38	0.75
I-Nova 2 Lite	Idatha yekhasimende kuphela	0.55	0.47
I-Nova 2 Lite	75% ikhasimende + 25% idatha ye-Nova	0.5	0.74
Qwen3-30B	Idatha yekhasimende kuphela	0.55	0.0038

Uma i-Nova 2 Lite icushwe kahle kusetshenziswa idatha yekhasimende kuphela, sibona a ukwehla okuphawulekayo kokunemba kwe-MMLU kusuka ku-0.75 kuya ku-0.47okubonisa ukulahlekelwa amakhono enhloso evamile. Ukucekelwa phansi kugqama nakakhulu kumodeli yakwaQwen, elahlekelwa kakhulu yikhono lokulandela imiyalelo ngemva kokucushwa kahle. Isibonelo sokukhishwa kwemodeli ye-Qwen eyehlisiwe:

{
  "prediction": "[n {n "content": "x^5 + 3x^3 + x^2 + 2x in Z_5",n "A": "0",n "B": "1",n "C": "0,1",n "D": "0,4",n "emotion": "neutral"n }n]"
}

Lokhu kuziphatha kuphinde kuhlobane nomklamo osheshayo we-VOC, lapho ulwazi lwesigaba lufakwa ngaphakathi ngokulungiswa kahle okugadiwe—indlela evamile ezinhlelweni zokuhlukanisa ngezigaba ezinkulu.

Ngokuphawulekayo, nini Ukuxuba idatha ye-Nova isetshenziswa ngesikhathi sokulungiswa kahle, i-Nova 2 Lite igcina ukusebenza okuvamile okuseduze kwesisekelo. Ukunemba kwe-MMLU kusasele 0.74kuphela 0.01 ngaphansi kwesisekelo sokuqala, kuyilapho i-VOC F1 isathuthuka 12 amaphuzu (0.38 → 0.50). Lokhu kuqinisekisa lokho Ukuxuba idatha ye-Nova kuyindlela esebenzayo nesebenzayo ukunciphisa ukukhohlwa okuyinhlekelele ngenkathi kugcinwa ukusebenza kwesizinda.

Okutholakele okubalulekile nezincomo ezisebenzayo

Lokhu kuhlola kubonisa ukuthi uma imodeli yesisekelo inikeza isisekelo esiqinile, ukuhlela kahle okugadwa ngokugcwele kwepharamitha ku-Amazon Nova Forge kungaletha izinzuzo ezinkulu zemisebenzi eyinkimbinkimbi yokuhlukanisa amabhizinisi. Ngesikhathi esifanayo, imiphumela iqinisekisa ukuthi ukukhohlwa okuyinhlekelele kuwukukhathazeka kwangempela ekukhiqizeni ukuhlela kahle ukuhamba komsebenzi. Ukucushwa kahle kwedatha yekhasimende kukodwa kungehlisa isithunzi amandla enhloso evamile njengomyalelo wokulandela nokucabanga, kukhawulele ukusetshenziswa kwemodeli kuzo zonke izimo zebhizinisi ezibanzi.

I Ikhono lokuhlanganisa idatha le-Nova Forge linikeza isu elisebenzayo lokunciphisa. Ngokuhlanganisa idatha yekhasimende namasethi edatha akhethiwe e-Nova phakathi nokulungisa kahle, amaqembu angakwazi ukulondoloza amakhono avamile aseduze nesisekelo kuyilapho eqhubeka nokuzuza ukusebenza okuqinile kwesizinda esithile.

Ngokusekelwe kulokhu okutholakele, sincoma le mikhuba elandelayo uma usebenzisa i-Nova Forge:

Sebenzisa ukulungisa okugadiwe ukuze ukhulise ukusebenza kwesizinda emisebenzini eyinkimbinkimbi noma eyenziwe ngendlela oyifisayo kakhulu.
Sebenzisa ukuxutshwa kwedatha ye-Nova lapho amamodeli kulindeleke ukuthi asekele ukuhamba komsebenzi kwenhloso evamile eminingi ekukhiqizweni, ukuze kwehliswe ingcuphe yokukhohlwa okuyinhlekelele.

Ngokuhlangene, lezi zinqubo zisiza ukulinganisela ukwenza imodeli ngendlela oyifisayo nokuqina kokukhiqiza, okuvumela ukusetshenziswa okuthembekile kwamamodeli ashunwe kahle ezindaweni zebhizinisi.

Isiphetho

Kulokhu okuthunyelwe, sibonise ukuthi izinhlangano zingakha kanjani amamodeli e-AI akhethekile ngaphandle kokudela ubuhlakani obujwayelekile ngamakhono okuxuba idatha we-Nova Forge. Ngokuya ngezimo zakho zokusebenzisa nezinjongo zebhizinisi, i-Nova Forge ingaletha ezinye izinzuzo, okuhlanganisa izindawo zokuhlola zokufinyelela kuzo zonke izigaba zokuthuthukiswa kwemodeli nokwenza ukufunda okuqinisiwe ngemisebenzi yomvuzo endaweni yangakini. Ukuze uqalise ngokuhlolwa kwakho, bona Umhlahlandlela Wonjiniyela we-Nova Forge ukuze uthole imibhalo enemininingwane.

Mayelana nababhali

I-Yuan Wei i-Applied Scientist kwa-Amazon Web Services, esebenza namakhasimende ebhizinisi ngobufakazi bomqondo kanye nokwelulekwa kobuchwepheshe. Usebenza ngokukhethekile kumamodeli ezilimi ezinkulu namamodeli olimi olubonwayo, agxile ekuhloleni amasu asafufusa ngaphansi kwedatha yomhlaba wangempela, izindleko, kanye nemikhawulo yesistimu.

Xin Hao unguchwepheshe Omkhulu we-AI/ML Go-to-Market kwa-AWS, esiza amakhasimende ukuba azuze impumelelo ngamamodeli e-Amazon Nova kanye nezixazululo ezihlobene ne-Generative AI. Unolwazi olunzulu lokusebenzisa i-cloud computing, i-AI/ML, ne-Generative AI. Ngaphambi kokujoyina i-AWS, u-Xin uchithe iminyaka engaphezu kweyi-10 emkhakheni wokukhiqiza izimboni, okuhlanganisa i-automation yezimboni kanye ne-CNC machining.

USharon Li i-AI/ML Specialist Solutions Architect e-Amazon Web Services (AWS) ezinze e-Boston, Massachusetts. Ngokuthanda ukusebenzisa ubuchwepheshe obusezingeni eliphezulu, uSharon uhamba phambili ekuthuthukiseni nasekuthumeleni izixazululo ze-AI ezikhiqizayo endaweni yesikhulumi samafu se-AWS.