Nhungamiro yehunyanzvi

Kuenderera mberi nePretraining vs Fine-Tuning

Kudzidzisa kwekufambidzira (continued pretraining) kunowedzera kudzidzisa modhi yekutanga ine chinangwa chifanana chekutarisa token inotevera pamusoro pemashoko makuru emuchidzidzo, nepo kudzidzisa kwekugadzirisa (fine-tuning) kuchidzidzisa pamusoro pemashoko madiki akasarudzwa ane zita, sekufanana kwekuraira nekudzokorora, kugadzirisa maitiro.

Teerera kugwaro iriAI-yakarondedzerwa · ~3 min

Pa peji ino4 min verenga
  1. Pfupiso
  2. Kudzika Kwakadzika
  3. Kuitwa muNyika yeChokwadi
  4. Zvakatipoteredza kweKudzidzisa Kwekufambidzira vs Kudzidzisa Kwekugadzirisa
  5. Kuzvishandisa
  6. Ramba Uchiongorora
  7. Mibvunzo inowanzo bvunzwa

Pfupiso

Sarudzo inokosha nekuti yekutanga inoshandura zvinangwa zve modhi nezve muchidzidzo nemutauro wacho, yechipiri inoshandura maitiro ekudzokorora, uye data nemagetsi anodiwa anokosha zvakanyanya.

Kudzika Kwakadzika

Continued pretraining, also called domain-adaptive pretraining, picks up where the original pretraining stopped. The model sees raw text from the target domain, such as medical papers, source code or legal filings, and learns to predict every next token, exactly as in pretraining. Gururangan and colleagues showed in the 2020 paper Don't Stop Pretraining that a second phase of pretraining on in-domain text improved RoBERTa on tasks in biomedical, computer science, news and review domains. It usually needs large volumes of text, often billions of tokens, and full-parameter training, so it costs far more than fine-tuning. Fine-tuning, usually supervised fine-tuning, uses prompt and response pairs. Loss is typically computed only on the response tokens, so the model learns what to produce given an input. Datasets range from a few thousand to hundreds of thousands of examples, and parameter-efficient methods such as LoRA make it affordable on modest hardware. It is good at teaching format, style, task behavior and instruction following. Preference methods such as RLHF or DPO often follow. A common pipeline is base model, then continued pretraining, then supervised fine-tuning, then preference tuning. Continued pretraining has a known risk: catastrophic forgetting, where general skills degrade. Applied to a chat model, it can also erode instruction following, so teams often start from a base model and redo instruction tuning afterward. Two misconceptions stand out. First, fine-tuning is not a reliable way to add many new facts; research such as Gekhman and colleagues (2024) found models learn new facts slowly through fine-tuning and that doing so can increase hallucination. For fast-changing facts, retrieval is often the better tool. Second, BloombergGPT is sometimes cited as continued pretraining, but it was trained from scratch on a mix of financial and general data. A rough rule: use continued pretraining when domain language differs a lot from web text and unlabeled data is plentiful; use fine-tuning when you need a behavior or format.

Kuitwa muNyika yeChokwadi

Code Llama yakagadzirwa nekufambidzira kudzidzisa Llama 2 pamusoro pemashoko makuru ecode, uye zvimwe zvezviratidzo zvayo zvakadzidziswa zvakare kwePython nekuteerera mairiziro.

Chikwata chekutevedzera mitemo ine zvikwata zvakawanda zvisina zita nezvikwata zvekutevedzera mitemo inoshandisa kudzidzisa kwekufambidzira kugadzirisa kushanda kwe modhi nemutauro wekutevedzera mitemo, uye inodzidzisa zvakare pamusoro pemashoko mashoma ekusarudza zvikwata zvinosarudzwa.

Bhanki inoda chatbot inodzokorora mune fomati yeJSON yakachengetedzwa nemutemo wakachengetedzwa; kudzidzisa kwekugadzirisa, kana kuti kuisa mibvunzo yakagadzirirwa, inokwana nekuti modhi inoziva zvinangwa zvepamutemo.

Meditron models yeEPFL yakagadzirisa Llama 2 kune medhichini nekudzidzisa kwekufambidzisa pamusoro pemabhuku e medhichini nemirairo yekurapa usati wadzidzisa basa rimwe nerimwe.

Zvakatipoteredza kweKudzidzisa Kwekufambidzira vs Kudzidzisa Kwekugadzirisa

Kuburitswa kwe modhi dzekutanga dzakakodzera dzakavhurwa kunogadzirisa kudzidzisa kwekufambidzisa kune zvikwata zvakawanda, nokunge zvinoda data nemagetsi makuru. Nzvimbo dzinofarira kufarira kusanganisa kutorwa, kugadzirisa muchidzidzo kugadzirisa, uye kusanganisa mabhenji kubva kune modhi dzakadzidziswa zvakasiyana. Mune maitiro, zvikwata zvakawanda zvinokanganisa kudzidzisa kwekugadzirisa kune retrieval pasi pekudzidzisa kwekufambidzira, zvichichengetedza nzira yakakosha kune muchidzidzo wemutauro usina kufarirwa nedata yepamutemo.

Kuzvishandisa

Nei zvichikosha

  • Mutengo uye bhajeti. Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore.
  • Sarudzo dzakajeka. Dzidzo yehunyanzvi inobatsira zvikwata kusarudza murwi wakakodzera, kwete iwo mutsva chete.
  • Kudzora kwemhando yepamusoro. Sarudzo dzeinjiniya dziri nani dzinoderedza zviitiko zvekuvimbika mukugadzira.

Chenjera

  • Kugadzirisa imwe bhenji kunogona kuvanza yakafara system kushaya simba.
  • Infrastructure uye mari yekugadzirisa inowanzotarisirwa pasi.
  • Chengetedzo uye kucherechedzwa mapundu anogona kukura sezvo masisitimu anowedzera kuoma.

Matanho anotevera

  1. Tsanangura latency, mhando, uye mutengo zvinangwa usati waitwa.
  2. Benchmark pasi pechokwadi mutoro uye data mamiriro.
  3. Chishandiso chekutarisa zvikanganiso, kudonha, uye mushandisi maitiro.
  4. Gadzirira nzira dzekudzosera kumashure uye dzezviitiko usati wawedzera.

Ramba Uchiongorora

Tsamba yemahara

Tora iyo yemazuva ese AI pfupiso

Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.

Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.

Zviedze iwe pachako

Tora quiz yeKudzidzisa Kwekufambidzira vs Kudzidzisa Kwekugadzirisa

Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.

Tanga mibvunzo

Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo

Mibvunzo inowanzo bvunzwa

Kudzidzisa Kwekufambidzira vs Kudzidzisa Kwekugadzirisa ndeipi?

Kudzidzisa kwekufambidzira kunowedzera kudzidzisa modhi yekutanga ine chinangwa chifanana chekutarisa token inotevera pamusoro pemashoko makuru emuchidzidzo, nepo kudzidzisa kwekugadzirisa kuchidzidzisa pamusoro pemashoko madiki akasarudzwa ane zita, sekufanana kwekuraira nekudzokorora, kugadzirisa maitiro. Sarudzo inokosha nekuti yekutanga inoshandura zvinangwa zve modhi nezve muchidzidzo nemutauro wacho, yechipiri inoshandura maitiro ekudzokorora, uye data nemagetsi anodiwa anokosha zvakanyanya.

Chii chisingafanirwe pakati pekudzidzisa kwekufambidzira nekudzidzisa kwekugadzirisa?

Kudzidzisa kwekufambidzira kunotarisa token inotevera pamusoro pemashoko, nepo kudzidzisa kwekugadzirisa kuchidzidzisa kubva kune prompt nekudzokorora.

Mune kudzidzisa kwekugadzirisa, panzi kusawirirana kunowanzoitwa?

Kumhanya prompt kunoreva kuti modhi inodzidza zvinangwa zvinodiwa kubva kune input, siri kutarisa input.

Replay ndeipi mune kudzidzisa kwekufambidzira?

Kuisa data yepamutemo kunobatsira kuchengetedza maitiro e pamutemo kuti isanyungudike nepo modhi ichidzidza muchidzidzo.

Code Llama yakagadziriswa sei?

Code Llama ndiyo inoratidza kudzidzisa kwekufambidzira, ine zvimwe zvezviratidzo zvakadzidziswa zvakare kwePython nemairiziro.

Kune zvinangwa zvinosiyana-siyana, nzira ipi inotsanangura kuti inokosha kupfuura kudzidzisa kwekugadzirisa?

Kugadzirisa kunodzidza facts itsva zvishoma uye kunogona kuwedzera kufungidzira, saka kutsvaga kunowanzova nani pakuchinja ruzivo.