Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.benchgen.com/llms.txt

Use this file to discover all available pages before exploring further.

Export Datasets → Train

The most direct path from an eval failure to a model improvement is exporting the failing cases as a training dataset and routing them straight into Train.

When to Export

Export a dataset when:
  • Your benchmark run shows a cluster of failures around a specific task or topic.
  • You have enough failing cases (rule of thumb: 50+ examples) to justify a fine-tune.
  • You’ve reviewed the failures and confirmed they’re fixable by training (not a prompt or scope issue).

How to Export

  1. Open the results report for the benchmark run you want to act on.
  2. Click Export → Send to Train.
  3. Choose which cases to include:
    • Failing cases only (recommended) — adds corrected labels automatically where available.
    • All cases — useful for building a balanced dataset.
  4. Name the dataset and confirm.
  5. The dataset appears immediately in Train → Datasets.

What Gets Exported

Each exported example includes:
  • The original input prompt
  • The model’s failing response (as a negative example, if applicable)
  • The expected/corrected response (as the training target)
  • Metadata tags from the benchmark (topic, error type)

Next Steps