SFT

Harbor includes utilities for turning trials (agent task completion attempts) into conversational traces that can be fed into supervised fine-tuning pipelines for agentic LLMs. Export helpers live under harbor.utils.traces_utils and power several CLI entry points.

• Each exported row represents one agent/episode-* directory and captures the input debug.json messages plus the final agent reply from response.json or response.txt.
• Rows include metadata such as agent, model, model_provider, task, trial_name, episode, and run_id, letting you merge runs from multiple jobs.
• --sharegpt adds a ShareGPT-style column to support instruction-tuning datasets expecting the {"from": "...", "value": "..."} schema.
• Success filtering (--filter success|failure) inspects result.json and lets you keep only passing or failing attempts for curriculum-style datasets.

Run harbor traces export on a trial directory (or a parent directory) to build a datasets.Dataset. The command prints the number of rows produced and, when --push is set, uploads directly to the Hugging Face Hub.

harbor traces export \
  --path trials \
  --recursive \
  --episodes last \
  --filter success \
  --sharegpt \
  --push \
  --repo my-org/harbor-terminus2-sft

Key options

If you want to persist the dataset locally (e.g., to Parquet), call the Python helper directly:

from harbor.utils.traces_utils import export_traces

dataset = export_traces("trials", episodes="last", success_filter="success")
dataset.to_parquet("harbor-terminus2-success.parquet")

The datasets library is an optional dependency; install it if you plan to export traces.

harbor run can export traces automatically once a job completes. Pass trace flags alongside your job invocation:

harbor run \
  --config examples/configs/job.yaml \
  --agent claude-code \
  --model anthropic/claude-3-sonnet-20240229 \
  --export-traces \
  --export-sharegpt \
  --export-episodes last \
  --export-push \
  --export-repo my-org/harbor-job-run

When --export-traces is set, Harbor exports from the produced job directory using the same machinery as harbor traces export. The --export-* options mirror the standalone CLI flags and default to in-memory exports unless --export-push is provided. Errors during export are surfaced at the end of the job run without interrupting evaluation.

harbor sweeps run can emit split datasets that separate successful and failed trajectories. Supply --push together with one of the repo arguments:

# Push a DatasetDict with "success" and "failure" splits
harbor sweeps run \
  --config examples/configs/job.yaml \
  --max-sweeps 3 \
  --trials-per-task 2 \
  --push \
  --export-repo my-org/harbor-sweeps

You can also push successes and failures to independent repos by combining --push with --export-separate (alias --no-export-splits) plus --export-repo-success and --export-repo-failure. These exports reuse the same trace discovery logic and default to the last episode from each trial.