Skip to main content
Traces are raw material for datasets. Any turn the connected agent handled (Connect an OpenClaw Agent) can become a dataset item; how to read those turns is on Benchmark an OpenClaw Agent.

​
Build a dataset from traces

On the Traces tab, select the turns you want (say, every benchgen-chat turn where the agent went wrong) and choose Add to dataset. Pick an existing dataset of this agent, or type a name (for example failed-turns) to create one.
Three chat turns selected on the Traces tab with the Add to dataset dialog open

Three chat turns selected, with Add to dataset open and a new dataset being created

Each selected trace becomes one dataset item with the turn’s input and output. The dataset is created in the agent’s own project and catalogued in BenchGen at the same time, so it:
  • appears under Datasets with its own URL, linking back to the traces it was built from;
  • can be picked in the training form, the same as any dataset uploaded by hand or imported from HuggingFace.
The Datasets page with a trace-built dataset carrying a Traces badge

The Datasets page: a dataset built from traces carries a Traces badge

The dataset's Data tab listing one item per trace

The dataset's Data tab: one item per trace, with its source trace, input and expected output

​
Find an item’s id and its source trace

The Data tab lists one row per item. The Item id column carries the id the training form and the API refer to, and Source trace links the item back to the trace it was made from; click it to open that trace in place, with the same observation tree and input/output view as on the Traces tab.
A dataset item's source trace opened from the Data tab

A dataset item's source trace opened straight from the Data tab

Items are not copied anywhere else: the dataset in the agent’s project is the one source of truth. Adding more traces later reports how many were added, skipped (already in the dataset) and failed. For human review before training, Add to annotation queue puts the same selection into one of the agent’s annotation queues.

​
Next: fine-tune a model

Add a Dataset

Or bring in more data by uploading a file or importing from HuggingFace.

Fine-tune a Model

Configure and launch a LoRA fine-tuning run on the dataset you just built.
Last modified on September 4, 2026