Lead Machine Learning Engineer | Generative AI | LLM | NLP

Arun Brahma  • Peerlist

Fine-tuned a Falcon-7B large language model using the QLoRA PEFT technique on a custom mental health conversational dataset. Achieved a training loss of 0.42 and achieved a ROUGE-1 score of 0.37 on the conversational test dataset.

Fine-tuning of LLM using QLoRA

Developed a prompt testing framework for LLM models to check prompt quality by adding multiple system prompts (using chain-of-thought, few-shot etc.), and then compare each system prompt based on the model-generated answer quality. Quality of answers can be measured using NLP metrics such as ROUGE, BLEU, or BERTScore and Responsible AI metrics such as Faithfulness, Answer Relevancy Score, Harmfulness etc.

Prompt Testing framework for LLM models

Transcribed spoken user queries using Automatic Speech Recognition model i.e. Whisper model and then generated relevant answers using large language models. The generated response is then passed to the TTS (text-to-speech) model for speech synthesis.

Smart Voice Assistant

Generated summarized text using OpenAI LLM for YouTube videos and built a conversational AI chatbot to get instant answers related to the extracted video transcript.