The latest AI model from Bhaiyabot, rohansolo/BB-L-01-7B, has been trained on 360k dialogue samples in Hindi, English, and Indian English. According to founder Rohan Shiralkar's post on LinkedIn, it outperforms Sarvam AI's OpenHathi-7B, Meta Llama-7B, and even Mistral AI's fine-tuned models while maintaining the performance of the base model. This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on HuggingFaceH4/ultrachat_200k and rohansolo/BB_HindiHinglish datasets. It achieves the following results on the evaluation set: Shiralkar states that AI in India is far behind and companies working in the AI field are too busy marketing non-achievements. This includes fine-tuning a model and labeling it as a pre-trained model, claiming to be India's first AI chatbot, without any product release or launch, such as BharatGPT - India's first AI in 14 Indian languages (LLM) - text, voice, video, and even fabricating facts like Krutrim. Shiralkar even questions Ola's recent launch of Krutrim. He writes on LinkedIn, "Ola's Krutrim claims to have trained a 2 trillion token LLM. And they have been active for 2 weeks. Is that even enough to train a mini model with 2 trillion tokens?" Furthermore, he says it is ridiculous that the news claims it is better than GPT-4. "I want an Indian LLM. I have cried for this for a long time. This is a strategic command. I just want a genuine one, not fabricated for fundraising," he adds.