लिखिए। आगे क्या आएगा?
Type a partial Hindi sentence. The model predicts what comes next — word by word, or the whole rest of the sentence.
Built for the messy middle of a sentence
Every design decision here traces back to a real fix made while training this model — nothing is decorative.
Fine-tuned, not from scratch
Starts from surajp/gpt2-hindi, an 88M-parameter GPT-2 pretrained on Hindi, then fine-tuned further on conversational Hindi lines pulled from OpenSubtitles.
Whole words, not subword scraps
GPT-2 predicts in subword pieces. The predictor walks token-by-token until it hits a real word boundary, so you see "कल्पना", never a stray "्पना".
Two modes, one model
Predict just the next word with ranked candidates, or let the model run on and stop itself after two or three full sentences.
Danda-aware stopping
Sentence completion counts Devanagari's । (danda) alongside ./?/! so generations end on a real sentence, not mid-token.