MultiSense/CustomerLM

CustomerLM: A Specialized Customer Simulator for Sales Benchmarking

CustomerLM is an 8 billion parameter large language model, fine-tuned from Qwen, specifically engineered to simulate the customer side of realistic sales conversations. Developed by MultiSense, its primary purpose is to serve as the user simulator for the SalesLLM benchmark.

Key Differentiator: Eliminating Role Inversion

Unlike general-purpose LLMs (e.g., GPT-4o) that tend to "assist" the salesperson when role-playing a buyer, CustomerLM is trained to maintain a consistent, in-character buyer persona. This significantly reduces "role inversion," where a simulated customer inadvertently helps the salesperson. CustomerLM cuts role inversion from 17.44% (GPT-4o) to 8.8%, producing responses measurably closer to real human customers.

Training and Capabilities

The model was trained in two stages (SFT then DPO) on 8,284 crowdworker-involved real-world sales dialogues in both Chinese and English. It learns to speak like a customer (SFT) and then specifically learns not to behave like a salesperson (DPO), using preference pairs that target this failure mode directly.

Good for:

• Driving the customer side of multi-turn sales dialogue simulations to benchmark or stress-test sales agents.
• Generating synthetic sales conversations for training or analysis.
• Role-play evaluation harnesses requiring a non-compliant, in-character human counterpart.

Important Considerations:

• Not an assistant: It is designed to act like a customer, asking questions and pushing back, not to be helpful to end-users.
• Format Sensitivity: Persona adherence is maximized when using the specified system prompt format for customer persona definition.
• Domain Coverage: Training data primarily covers Financial Services and Consumer Goods.