MultiSense/CustomerLM

CustomerLM: A Specialized Customer Simulator

CustomerLM is an 8 billion parameter language model, fine-tuned from Qwen, engineered to act as a realistic customer in sales conversation simulations. Unlike general-purpose LLMs that tend to be overly helpful, CustomerLM is trained to maintain a skeptical, self-interested, and non-compliant buyer persona, significantly reducing "role inversion" where the customer acts like a salesperson.

Key Capabilities

• Realistic Customer Simulation: Trained on 8,284 crowdworker-involved real-world sales dialogues, it accurately mimics human customer behavior.
• Reduced Role Inversion: Achieves an 8.8% role inversion rate, a 2x reduction compared to GPT-4o's 17.44%, ensuring more reliable sales agent evaluations.
• Persona Adherence: Stays in character as a buyer, including being terse, skeptical, and willing to walk away, based on a detailed system prompt.
• Multilingual Support: Supports both Chinese and English sales dialogues.

How it Works

CustomerLM uses a unique role mapping where the system prompt defines the customer persona, user messages are the salesperson's turns, and assistant replies are the customer's responses. It was trained in two stages: Supervised Fine-Tuning (SFT) to learn customer-like speech, followed by Direct Preference Optimization (DPO) to specifically eliminate salesperson-like behavior.

Good For

• Driving the customer side of multi-turn sales dialogue simulations for benchmarking or stress-testing sales agents.
• Generating synthetic sales conversations for training, analysis, or research.
• Any role-play evaluation harness requiring a non-compliant, in-character human counterpart.

Do not use it as an end-user assistant; it is designed to push back and ask questions, not to be helpful or provide factual information.