Selling-Assistant-V1

Introduction

Selling-Assistant-V1 is a state-of-the-art Sales Language Model built upon the Qwen3-30B-A3B architecture. Unlike complex agentic systems with separate classification modules, Selling-Assistant-V1 is an end-to-end generative model optimized via Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) to master the art of persuasion, negotiation, and customer service.

It internalizes complex sales logic—from rapport building to closing deals—directly into its parameters, offering a streamlined, high-performance solution for e-commerce and CRM applications.

Performances on Benchmarks

We evaluate Selling-Assistant-V1 against leading general-purpose models on a proprietary Sales Capability Benchmark, which assesses performance across four critical dimensions: Persuasion Rate, Empathy Score, Objection Handling, and Compliance.

Benchmark (Sales Domain)	Selling-Assistant-V1	Qwen3-30B-A3B	Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT
Persuasion Rate	85.4%	72.1%	78.5%
Empathy Score (0-10)	9.2	7.8	8.1
Objection Handling	88.9%	75.4%	79.2%
Rule Compliance	99.1%	85.0%	88.5%
CSAT Proxy	4.8/5	4.2/5	4.4/5

Evaluation Parameters

Default Settings (Sales Tasks)

• Temperature: 0.7
• Top-p: 0.9
• Max new tokens: 512
• System Prompt: Standard Sales Assistant Persona

For Objection Handling scenarios, we utilize a lower temperature (0.5) to ensure consistency and adherence to approved counter-arguments.

Core Sales Techniques

The model has been rigorously trained on top-tier sales methodologies, enabling it to naturally exhibit the following behaviors without external prompting:

1. Trust Establishment & Needs Discovery

   • SPIN Questioning: Naturally sequences Situation, Problem, Implication, and Need-payoff questions.
   • Empathetic Resonance: Validates customer emotions before proposing solutions.

2. Value Alignment

   • FABE Framework: Automatically translates product Features into Customer Benefits.

3. Deal Acceleration

   • Objection Neutralization: Addresses pricing and quality concerns with proven scripts.
   • Closing Strategies: Identifies buying signals and applies soft closes.

4. Retention & Growth

   • Cross-Selling: Contextually suggests relevant add-ons (Upsell/Cross-sell).

Serve Selling-Assistant-V1 Locally

For local deployment, Selling-Assistant-V1 supports high-performance inference frameworks including vLLM and SGLang.

vLLM

Install vLLM (ensure compatibility with your CUDA version):

pip install -U vllm

Start the server:

vllm serve digitalassistant-ai/Selling-Assistant-V1 \
     --tensor-parallel-size 1 \
     --gpu-memory-utilization 0.95 \
     --max-model-len 32768 \
     --served-model-name selling-assistant-v1

SGLang

Install SGLang:

pip install "sglang[all]"

Launch the server:

python3 -m sglang.launch_server \
  --model-path digitalassistant-ai/Selling-Assistant-V1 \
  --tp-size 1 \
  --port 8000 \
  --host 0.0.0.0 \
  --served-model-name selling-assistant-v1

Transformers

Basic inference using Hugging Face Transformers:

pip install transformers accelerate torch

Python Code:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL_PATH = "digitalassistant-ai/Selling-Assistant-V1"

tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH, 
    device_map="auto", 
    torch_dtype="auto"
)

messages = [
    {"role": "system", "content": "You are a professional sales assistant."},
    {"role": "user", "content": "This phone is too expensive."}
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    inputs, 
    max_new_tokens=256, 
    temperature=0.7, 
    top_p=0.9
)

print(tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True))

Future Roadmap: Agentic Architecture

While V1 focuses on end-to-end generation, our next-generation Selling-Agent-V2 will evolve into a fully autonomous system leveraging the Model Context Protocol (MCP). This architecture separates cognitive reasoning from tool execution, enabling deeper integration with enterprise ecosystems.

Architectural Highlights

1. Standardization of MCP Protocol

   • Unified Tool Interface: A central MCP Protocol Hub standardizes interactions with external tools, allowing the agent to seamlessly query CRM data, calculate complex discounts, and access real-time competitive intelligence.
   • Enterprise Integration: Direct synchronization with sales opportunity statuses and dynamic retrieval of product parameters via FAB libraries.

2. Advanced Agent Brain

   • Dual-Layer Memory: Combines Short-term Session Context with Long-term Customer Profiles (CDP) to deliver hyper-personalized interactions across multiple touchpoints.
   • Strategic Planning: Implements Sales SOP Path Planning and Chain-of-Thought (CoT) reasoning to navigate complex negotiations and perform self-censorship for compliance.

3. Modular Core Components

   • The architecture re-integrates specialized modules for Intent Recognition, Sentiment Analysis, Conversion Prediction, and Quality Assurance, providing granular control and explainability over the sales process.

This evolution marks the transition from simulating a salesperson to deploying an autonomous, tool-augmented sales employee.

License

This project is licensed under the Apache 2.0 License.

Citation

If you use this model in your research or application, please cite:

@misc{selling_assistant_v1,
  author = {Selling AI Team},
  title = {Selling-Assistant-V1: A Specialized Chinese Sales Language Model},
  year = {2024},
  publisher = {Hugging Face},
  journal = {Hugging Face Repository},
  howpublished = {\url{https://huggingface.co/digitalassistant-ai/Selling-Assistant-V1}}
}