(Emergent Abilities) aagyi
Abe lol isse hame other language me less trian kiya tha tab bhi ye wo language kaise sikh gya ? How lol ?
Kuch samjh ni aarha harshit batna na
@shrijanagain @gkrideai
Fast team se bro aur shi krwo samjhe me shards ko separate krne ke badd baat krunga
How lol load kiye model
Ha karwe 3 times same chiz notice kiya tab yaha par aaya hu
Lag to muje bi rha tha ye chiz ya other languages ke data collection 0 percent the tabhi sikh gyaa
Wohi tab acesss off post maro ye discussion me
"Bhai 'how' puchne ka time nahi hai. 'Why' baad mein dekhenge. Abhi 'stop' karna hai. Yeh emergent behavior hai - model ne apne aap seekh liya jo humne nahi diya. Iska matlab internal representations unpredictable hain. Agar languages automatic sikh sakta hai, toh aur kya sikh sakta hai?"
Today I learned that Hindglish was a thing, was a fun rabbit hole to dig into.
That said, models extrapolating (semi decent) language understanding from very limited datasets (or even datasets containing only a few words here and there in another language) is nothing new, it's even precisely what LLM (language model, it's even in the name 😄) are good at. Plenty of LLM certified for English only will still be able to speak some french, Japanese, and so on. If your training data is as large as you say it is, the opposite would actually be the surprising part.
Anyway, have fun, and good luck.
Hehe, true! 🌟 LLMs sach mein magic jaise hote hain na? Par 'bol paane' aur 'dil se samajhne' mein thoda sa fark hota hai.
Hinglish sirf words mix karna nahi, ek vibe hai. Humne ShORT-Hinglish ke 1.1 Trillion tokens ke sath wahi vibe pakadne ki koshish ki hai, taaki ye sirf 'extrapolate' na kare, balki native lage. 💖
Anyway, thanks for the wishes! ✨ Aapka logic bhi sahi hai, par benchmarks ke baad shayad aapko thoda aur surprise milega. Have fun too! 🤞