Domain Fine-Tuning with Qwen for Business-Specific Sales Analysis

#38
by ahmadzainuri - opened

Hi everyone,

I’m trying to understand the best way to approach domain fine-tuning with Qwen.

Let’s say I have a general-purpose Qwen model, and I want to make it better at understanding a specific business domain, for example sales/retail.

The general model already understands what “sales” means, but I want to teach it a company-specific way of analyzing sales performance.

General Knowledge

For example, the model already knows that:

  • Revenue = price × quantity
  • Sales can be analyzed by product, region, salesperson, etc.
  • Higher sales generally indicate better performance.

Company-Specific / Domain Knowledge

However, our company evaluates sales performance using these 5 factors:

  1. Revenue achievement
  2. Quantity achievement
  3. Gross margin
  4. Customer retention
  5. Sales conversion rate

Our company also has a specific methodology for interpreting these factors.

For example:

Question

Why did Salesperson A have lower sales performance this month?

Expected Domain-Specific Answer

Salesperson A's performance should not be evaluated only based on revenue. According to our company's methodology, we should also consider quantity achievement, gross margin, customer retention, and conversion rate.

For example, if revenue decreased by 10%, but gross margin and customer retention increased, the performance should not automatically be considered worse.

My Questions

  1. Is this a good use case for domain fine-tuning / SFT?
  2. Would you recommend LoRA or QLoRA for this type of training?
  3. Should the training dataset consist mainly of Question → Expert Answer examples like the example above?
  4. How many high-quality examples would typically be needed to teach this kind of domain-specific behavior?
  5. Would fine-tuning actually teach the model this company's way of reasoning, or would RAG be more appropriate?
  6. If I later add company documents and real-time sales data through RAG/database, would the recommended architecture be: Qwen + Domain Fine-Tuning + RAG + Database ?

I’m particularly interested in hearing from people who have fine-tuned Qwen for a specific business/domain, rather than just general instruction tuning.

Thanks!

Sign up or log in to comment