Stage - Data Science NLP H/F
BNP Paribas
StageSur site4 à 6 moisRémunéréDate limite : 6 nov. 2025
Data Science & IAnatural language processingAI Agents & Autonomous Systems
Description
- The NLP team within BNP Paribas Cardif's Corporate Analytics is actively exploring the latest advancements in Large Language Models (LLMs) to enhance their reliability within our conversational AI systems.
- These systems serve various use cases, such as assisting internal employees, educational chatbots, and customer support assistance to improve customer experience and operational efficiency.
- While Retrieval Augmented Generation (RAG) has improved access to corporate knowledge, hallucinations (confident but inaccurate or unverifiable responses) remain a key obstacle to deploying such models responsibly in real financial contexts.
- This internship aims to develop robust and explainable RAG architectures that mitigate hallucinations, increase response transparency, and align model behavior with Cardif's trust, security, and reliability standards.
- Responsibilities include:
- Conducting a literature review on hallucination detection, response abstention capability, and uncertainty estimation of LLMs.
- Designing explainable RAG pipelines capable of providing source attributions or citations for generated responses.
- Integrating human or synthetic feedback using alignment techniques such as Direct Preference Optimization (DPO) or Reinforcement Learning from Human Feedback (RLHF).
- Studying dense conversational retrievers that operate directly on the conversation history without explicit query rewriting to enhance retrieval performance in multi-turn conversations.
- Developing and evaluating agent-based RAG systems capable of reasoning, retrieving, and verifying information dynamically throughout the conversation.
- Evaluating confidence measures and hallucination scores for online and offline evaluation metrics.