Sapient Intelligence
ShowcaseDashboard
Invite
Sapient Intelligence
Showcase
  • Banking Support Intent Recognition
  • Customer Review Analysis
  • Enterprise Tool Calling
  • Vehicle Owner Feedback Classification
  • Natural Language Data Query
Dashboard

Banking Support Intent Recognition

Understand what the customer is really asking for

A banking support message could be about transfers, fees, card issues, or many other detailed requests. This shows how the system learns to tell apart 77 types of banking support intents.

Agent task replay

How the agent creates the model

Uploaded train.jsonl

Uploaded check dataset inputs/train.jsonl (9988 rows). Validated structure successfully.

Uploaded test.jsonl

Uploaded check dataset inputs/test.jsonl (3076 rows). Validated structure successfully.

Parameter configured

Configured parameters for the fine-tuning run.

Training plan confirmed

Generated task.yaml. User confirmed task parameters and triggered the execution pipeline.

Fine-tuning executed

Fine-tuning complete. Evaluated macro_f1 score is 0.9298.

We get thousands of customer messages a day and my team hand-routes each one to the right desk. I want something that reads a message and tells me which of our 77 request types it is. I already have a big set of past messages, each already labeled with the right request type, plus a separate set to check it on.
user
After adding label boundaries, negative samples, or evaluating preferences, the agent will include them in the current mock run.

Result Comparison

See if it actually gets the customer request right

Test samples show what changed before and after training, especially whether the system stops mixing up similar requests like cash withdrawal fees and cash withdrawals.

Key effect: Key metric comparison

The fine-tuned model learns the task boundary from the provided examples and fixes common label confusions.

0.9298
+0.650

Fine-tuned macro_f1

Primary metric · n=3076

0.2793

Base macro_f1

Zero-shot HRM-XL-1.2B

92.98%
+65.1 pts

Fine-tuned Accuracy

Verified prediction accuracy

0 / 3,076
was 663

OOV Predictions

Closed label compliance

Test set probe: Test-set sample probe

Input
Base Model
Fine-tuned
Correct
Status
View

Selected test case

Worse

Customer message: How do I locate my card?

Basecard_arrival
Fine-tunedlost_or_stolen_card
Goldcard_arrival

Private alpha

Create your first exclusive model

Users with an invitation code can proceed to the Agent guidance process. Those without an invitation code can still submit their requirements and join the Waitlist.

Try to createJoin the Waitlist