Post-Training SeeWhy: A Small Specialized Model with Intuit/Credit Karma.

Together with Intuit Credit Karma, we post-trained SeeWhy, a small, specialized model that explains why a recommended credit card fits a particular user. This collaboration focuses on personalized explanations that preserve compliance-critical details, including card names, fees, and interest rates.

The post-training approach

The pipeline combines synthetic data with engineered distributions, supervised fine-tuning of Llama 3.1 8B with LoRA, and GEPA prompt optimization. Teacher-generated explanations undergo rejection sampling before training. Evaluations then check product accuracy, legal and partner compliance, and product requirements.

A central idea is Dynamic Semantic Tags (DSTs). These semantic XML tags associate each critical value with its product and attribute, helping the model copy the correct source values. Assigned during prompt construction, the tags also allow generated values to be checked programmatically against the underlying card data.

Offline results

In offline simulations, the fine-tuned model achieves 100% card-name accuracy, 99.2% fee accuracy, and 99.9% rate accuracy. Average latency falls from 3.41 to 3.06 seconds compared with gpt-4.1-mini. The report also describes 15–20× lower serving cost through self-hosting.

Read the full technical report: Dynamic Semantic Tags Reduce Hallucinations in Small-LLM Post-Training.