kaman-asan-chatbot

Chat with Hooman

A small Banglish chatbot. You write a message in standard Banglish or in Hooman Dialect; it replies in Hooman Dialect.

kaman asan           ->  Ame valo asi, apnar ki kobor
kemon achen          ->  Alhamdulillah valo asi, apnar ki kobor
bhat khaisen         ->  Ame to kai, sobsomoi kayacan
apnar basha kothay   ->  Basa nae, rastai taki
apni ki manush       ->  Ame manus, apna ki
dhonnobad            ->  Oilcome

Try it: Chat with Hooman

What it is

google/byt5-small (300M), fine-tuned in two stages:

1. a transliteration stage on ~2k pairs mapping standard Banglish to Hooman Dialect, which taught it the spelling system
2. a dialogue stage on ~900 message→reply turns, which taught it to answer rather than translate

Byte-level tokenisation matters here: Banglish has no standard spelling, so khaisen, khaichen and kaisen are all the same word, and a subword tokeniser splits them inconsistently.

Use

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("chaos24/kaman-asan-chatbot")
net = AutoModelForSeq2SeqLM.from_pretrained("chaos24/kaman-asan-chatbot")

enc = tok("kaman asan", return_tensors="pt")
out = net.generate(**enc, do_sample=True, temperature=0.8, top_p=0.92, max_new_tokens=128)
print(tok.decode(out[0], skip_special_tokens=True))

Sample rather than decode greedily — greedy gives the same answer every time, which reads like a lookup table instead of a conversation.

Limitations

• Single-turn. It sees only your latest message and has no memory of the conversation.
• ~900 training turns. It handles the topics it was trained on — greetings, food, sleep, where you live, playful teasing — and gets vague or wrong outside them. That is the ceiling of the data, not something more training fixes.
• Validation chrF is 52.96. For open-ended dialogue that mostly measures overlap with one reference reply out of many valid ones, so read it as a training signal, not as quality.
• No safety tuning of any kind. It is a toy.

Hooman Dialect

A language made up for this project — not Noakhali, Sylheti, Chittagonian, or any real region's speech, and not a claim about how anyone talks. Roughly: aspirates lose their h (bhat→bat), j→g (kaj→kag), e→a (achen→asan), and ai→aya inside a word (khaisen→kayasan).

Training data came from a private conversation with names and locations removed.