Prove the model improved. See where it didn’t.
Compare your base model and fine-tuned adapter on the same held-out examples. Measure the task you care about, then inspect the answers behind the score.
Compare generated answers against references, with metrics such as Token F1.
Interactive illustration · sample data
One test set. A clearer comparison.. Explore illustrative results for different kinds of model output.
Go beyond a single score.
A useful evaluation tells you what changed, which examples regressed, and whether your test data supports the conclusion.
Compare on equal terms
The fine-tuned evaluation reuses its baseline’s test data, scoring, and generation settings. Pair the answers example by example to see the effect of training.
Explore the featureMatch the metric to the task
Use Q&A, classification, structured JSON, or numeric presets. Add an LLM judge with your rubric, or register custom Python metrics.
Explore the featureFind the regressions
Inspect improved, unchanged, and worse answers. Break results down by category, recipe, or source to see what an average can hide.
Explore the featureUnderstand the uncertainty
See a likely range for the score change when there are enough independent test documents. When there is not enough evidence, the comparison says why.
Explore the featureA better average is only the beginning.
Read the answers, look at the failures, and decide what to improve in your data. Bring that insight into the next version instead of treating evaluation as a final checkbox.
- Held-out examples
- Score on test data that training never sees.
- Your model endpoint
- Evaluate a ready deployment or your own connected endpoint.
- Repeatable decisions
- Use experiment acceptance thresholds and CLI quality gates.
Build with Tensorant