Hinglish is a hybrid language blending Hindi and English, commonly spoken in India, combining vocabulary and grammar from both languages. Hinglish is often used in text conversations by people in India. Hinglish text mostly contains english characters, transliterated from hindi sentences. Example: "Aaj ka din bohot acha hai".
It is helpful to have Hinglish datasets to finetune opensource LLMs like LLAMA-2 that haven't seen such data in the training phase. GPT-3 and onwards model however, have already seen hinglish data during training.
Since GPT-3 and above have seen hinglish data during training, we leverage them to generate conversations, which are further post-processed to generate a clean dataset. GPT prompt used was:
I want you to generate a Hinglish conversation between two young Indians - a male and a female. Feel free to assume the names of these young Indians. The conversation should contain 100 dialogues. Conversation should be in the format [Name]: [Message]. Conversation should be strictly in Hinglish. If the conversation happens in English, I will punish you. The conversation should be slightly flirty in nature - ending in a romantic moment. The conversation is around the topic: '{setting}'. Do not change subjects frequently. If possible, talk about a subject at length.
See topics.md for the various conversation contexts provided in the prompt.
The data is post-processed so that each line in the conversations is strictly in the format [Name]: [Dialogue]
Feel free to add more conversations!