AI Training

The AI Training page lets you collect, curate, and export training data pairs for fine-tuning language models.

Sidebar item: AI Training Route: /training

What Are Training Pairs?

A training pair consists of a user prompt and a model response. By collecting high-quality pairs, you build a dataset that can be used to fine-tune a model to behave the way you want.

How Pairs Are Collected

Training pairs are automatically captured from your interactions with the Canvas AI feature in the Tool Workshop. When the AI generates tool code, drafts improvements, or provides insights, each interaction becomes a training pair stored in IndexedDB.

Curating Data

The main interface shows a table of all collected training pairs. For each pair you can:

Accept or Reject

• Accept — Mark the pair as high-quality training data
• Reject — Mark it as unsuitable (bad output, hallucination, etc.)

Edit Responses

Click on any response to edit it. This is useful when the model's output was close but needs correction — you get the right answer without starting from scratch.

Tagging

Add tags to organize pairs by topic, quality level, or any custom category. Tags help you filter and export specific subsets.

Bulk Operations

Select multiple pairs to:

• Bulk accept/reject — Set status for many pairs at once
• Bulk add tags — Apply a tag to all selected pairs
• Delete — Remove unwanted pairs

Filtering

Filter the training pair list by:

• Status — Accepted, rejected, or unreviewed
• Tags — Show only pairs with specific tags
• Search — Free-text search across prompts and responses

Exporting

Select the pairs you want to export (or use filters to narrow down), then click Export. The exported dataset can be used with fine-tuning tools and frameworks.

Statistics

The page header shows aggregate stats:

• Total pairs collected
• Accepted vs. rejected counts
• Tag distribution

Tips

• Quality over quantity — A small dataset of carefully curated pairs produces better fine-tuning results than a large noisy dataset.
• Edit, don't discard — If a response is 80% correct, editing it is more efficient than regenerating from scratch.
• Use tags strategically — Tag by capability (e.g., "code-gen", "api-tools", "explanation") to create focused training subsets.