Lesson 145 — Advanced

Should You Fine-Tune?

Fine-tuning sabse exciting option lagta hai, par ye sabse mehnga aur sabse zyada long-term commitment wala option hai. Is lesson mein hum option ladder dekhenge — prompt se fine-tune tak —, fine-tune kis kaam ke liye sahi hai aur kis ke liye nahi, aur training shuru karne se pehle poora kharcha kaise naapa jaata hai.

⏱ 10 min• Advanced• Basic computer usage

Fine-tune — sabse kharchila rung

Kya hai?

Fine-tuning ka matlab hai base model ke weights ko apne hi data se dobara train karke usse apne kaam ke liye special banana. Prompt se behaviour maangne ki jagah, behaviour ko weights mein baitha dete ho — model ko sunane ke bajaye uski aadat badal dete ho.

Kab kaam aata hai?

Jab prompt, retrieval aur tools ke baavjood ek hi measured failure baar baar dohraye — tone consistent nahi baith rahi, output format prompt se pakda nahi ja raha, ya narrow task pe har request pe lamba prompt dena pad raha ho.

Kahan milta hai?

Managed fine-tuning APIs mein, ya apni GPU pe training scripts mein (TRL, PEFT, Axolotl, Unsloth jaisi tooling). Par sirf training script nahi — data pipeline, eval suite aur serving layer teeno chahiye.

Funda ye hai: prompt tumhe control deta hai, fine-tune tumhe habit deta hai. Control badalna sasta hai (kal prompt likh diya, parso badal diya); habit badalna mehnga — data, training, eval aur phir base model ke saath uski poori zindagi ka kharcha.

Rule of thumb: Jab tak tum ek measured failure ka naam likh nahi sakte — kaunsi task, kitni buri, kaunsa metric — fine-tune mat shuru karo. Andhere mein train karna sirf mehnga experiment hai, result ka koi paimaana nahi hoga.

Option ladder — prompt se fine-tune tak

0. clear prompt + few examples   -> sabse sasta, turant badal sakte ho
1. retrieval (docs context)      -> facts jo badalte hain, model ko yaad nahi karane
2. tools / structured output     -> kaam external system se karana (DB, API)
3. stronger base model           -> capability hi chhoti hai, behaviour ki nahi
4. fine-tune                     -> behaviour/format baithana — sabse mehnga, sabse last

har rung tabhi chadho jab neeche wala rung measurably fail ho chuka ho

0 — Prompt: agar ek achhi instruction se problem hal hoti hai, yahi karo. Cost zero, rollback zero, kal badal sakte ho.
1 — Retrieval: jawab ko bahar ka fresh data chahiye (price, policy, inventory) toh model ko yaad mat karwao, context do.
2 — Tools: model ko compute, lookup ya action chahiye — fine-tune uska substitute nahi, tool hi lagta hai.
3 — Stronger model: tumhara kaam bahut general hai (writing, reasoning) toh chhota model ko tune karne se behtar hai behtar base le lena.
4 — Fine-tune: tabhi, jab behaviour ki inconsistency ya format ka failure numberon mein dikha ho.

Ladder ka matlab yeh nahi ki sab chadhna hai: Aadhi teams pehla hi rung chhod kar seedha fine-tune pe kood jaati hain. Ek hafta prompt aur retrieval behtar karke dekho — aksar wahi gap band ho jaata hai, aur bacha hua waqt hi asli saving hai.

Fine-tune kis kaam ke liye sahi hai

Consistent style aur tone: har support reply ek hi brand voice mein ho — prompt se kabhi kabhi chhoot jaati hai, weights se nahi.
Formats jo prompt hold nahi kar paata: lamba, strict structured output jo har response mein bilkul same shape mein chahiye.
Narrow structured tasks: classification, extraction, routing — ek chhota kaam jo hazaar baar dohrata hai.
Chhote prompts: tune ke baad 2000-token system prompt hata kar chhota kar sakte ho — latency aur token dono bachte hain.
Chhota specialised model: bade teacher-model ki aadat chhote student-model mein baithana, taaki khud host kar sako.

FINE-TUNE SE SOLVE HOTA HAI        RETRIEVAL / TOOLS SE SOLVE HOTA HAI
consistent tone & style            fresh facts (price, policy, stock)
fixed output shape                 user/account specific data
narrow repeated task                "abhi wali" jaankari
chhota prompt chalana               task jiska definition hi badalta rehta hai

yaad rakho: model ko "kya pata hai" tune se nahi,
"kaise behave karta hai" tune se badalta hai

Task hi define nahi hua toh training andheri hai: Agar ye tay nahi kiya gaya ki "accha jawab" hai kya, toh tune karke kya prove karoge? Eval banane layak task pehle likho — jin tasks ko naap hi nahi sakte unpe fine-tune sabse kharab investment hai.

Poora cost — sirf GPU bill nahi

COST OF OWNING A TUNE
1. data creation      -> examples likhna/saaf karna/version karna (aksar sabse mehnga)
2. training runs      -> ek run nahi, hyperparameter ke kai experiments
3. evaluation         -> baseline vs tuned, held-out set, regressions check
4. serving            -> tuned weights host karna, template/tokenizer match
5. retraining         -> naya base model aaya toh data + train + eval dobara

compare this with: ek prompt edit jo turant lag jaata hai

Data ek baar nahi banta: har naye failure case ke saath dataset badalta hai — cleaning aur versioning ongoing kaam hai.
Training ek run nahi hota: curve padho, checkpoint chuno, dobara chalao — compute ka kharcha multiply hota hai.
Serving bhi free nahi: tuned model ko serve karna, monitor karna aur rollback rakhna — operations ka kharcha.
Base model ke saath tied: fine-tune sirf usi base version pe chalega — naya base aaya toh poori journey dobara.

Cost comparison ka sahi tarika: Prompt-wale option ka kharcha (tokens per request × traffic) bhi side-by-side likho. Bahut baar tuned model ka token saving hi tune ka pura kharcha chukata hai — par ye number nikalne ke liye dono taraf ka hisaab chahiye, andaza nahi.

Exercise

Question: Fine-tune karne se pehle option ladder mein sabse pehle kaunsa option try karte hain?

Question: Jo facts regularly badalte hain (price, policy) — unhe fine-tune ki jagah kis cheez se supply karte hain?

Question: Fine-tuning se model ka behaviour kahan baith jaata hai — prompt mein ya weights mein?

Question: Naya base model aa gaya — purana fine-tune uspe badalne ka naam kya hai: sirf model name change ya dobara training?

Common mistakes

Measured failure ke bina shuru karna: Sirf "fine-tune se behtar hoga" kehte hue train kar diya, par pehle ka failure number mein likha hi nahi. Fix: task, current failure rate aur target metric — teeno likho, tabhi training shuru.
Changing facts sikhane ke liye fine-tune karna: Policy ya price weights mein daal di toh har update pe dobara train karna padega. Fix: facts retrieval se, behaviour (style/format) weights se.
Sirf GPU ka bill gin na: Data creation, eval aur serving asli kharcha hain; GPU toh chhota hissa hai. Fix: poora ownership cost likho — data + runs + eval + serve + retrain.
Task ka eval hi na hona: Naapne layak task nahi hai toh tune ke baad "behtar laga" se zyada kuch nahi bachega. Fix: pehle eval set banao, phir training ka faisla lo.

Desi example: Ready-made suit vs tailor ka order

Ready-made suit pehni nahi toh tailor ke paas gaye — par tailor ka order mehnga, dheere aur har naye season mein silwana padta hai. Pehle ready-made ki fitting theek karo (prompt), andar ka support badlo (retrieval), tabhi measurably fail ho toh tailor ko bolo (fine-tune):

ready-made   -> prompt / examples   (sasta, palat sakte ho)
fitting      -> retrieval / tools   (bahar ka maal andar)
custom tailor-> fine-tune           (mehnga, ek hi base pe sila)
naya fashion -> naya base model     -> tailor ka kaam dobara

Interview angle — yahan se sawaal aate hain

Q: Fine-tune aur RAG ka farq kya hai? — RAG facts deta hai jo badalte hain; fine-tune behaviour badalta hai jo weights mein baithta hai.
Q: Tumhe kaise pata chalega ki tune ki zarurat hai? — Prompt/retrieval ke baad bhi ek named, measured failure bacha ho, jiska apna metric ho.
Q: Fine-tune ka sabse bada long-term kharcha kya hai? — Ek hi base model version se tie, isliye naya base aane pe poora kaam dobara.
Q: Cost comparison kaise karoge? — Prompt tokens per request vs tuned serving cost, plus data/train/eval/serve ka ownership cost side-by-side.
Q: Kaunse task ke liye fine-tune galat choice hai? — Jiska definition/eval hi nahi bana, ya jismein jaankari roz badalti hai.

Lesson complete?

Decision framework ready — ab tum reasoned tarkee se bologe ki tune karna hai ya nahi. Agla lesson — Data Curation for Fine-Tuning — mein dekhenge us decision ke baad data kaise banta hai.