Output Format

Tracer archives rollout outputs under artifacts/. The primary payload is a set of Harbor jobs: each job contains per-task trajectories, aggregate rollout results, optional SFT conversion output, and run history for auditing.

Artifact layout

The main Tracer outputs are organized as follows:

artifacts/
├── tasks/                              # prepared task directories staged for Harbor
│   └── <dataset>/                      # local or Hugging Face task source
├── jobs/                               # Harbor rollout jobs
│   └── <job>/
│       ├── result.json                 # aggregate rewards, errors, and job stats
│       ├── config.yaml                 # active config snapshot at launch
│       └── <task>/
│           └── agent/
│               └── litellm-trajectory.jsonl # replayable trajectory for one task
├── sft_data/                           # optional converted training data
│   └── <job>/
│       ├── im.jsonl                    # intermediate records with quality metadata
│       ├── lf.json                     # LLaMA-Factory ShareGPT data
│       ├── lf.stats.json               # token, turn, and score statistics
│       └── .convert_sig.json           # conversion cache signature
├── litellm/                            # per-job LiteLLM proxy configs
├── logs/                               # rollout, proxy, and launch logs
├── processed_tasks.yaml                # task IDs already consumed by Tracer
├── archives/                           # immutable run snapshots
└── index.yaml                          # append-only run history

Most users should start with artifacts/jobs/<job>/, then inspect artifacts/sft_data/<job>/ only when training-data conversion is enabled. The dashboard reads the same artifact tree, so these files are also the easiest way to debug a run without learning the internal scripts first.

Task preparation outputs

Before rollout, Tracer stages compatible task directories into:

artifacts/tasks/<dataset>/

These tasks come from Curator output or a Hugging Face source. A Curator source is staged as symlinks back into its pool, so this directory lists the tasks in play without holding a second copy of them. The processed task ledger is stored at:

artifacts/processed_tasks.yaml

Use the ledger to avoid rerunning tasks that were already consumed. It is runtime state, not a hand-written task list.

Rollout job outputs

Each rollout job writes one directory under artifacts/jobs/:

artifacts/jobs/<job>/
├── result.json
├── config.yaml
└── <task>/
    ├── result.json
    └── agent/
        └── litellm-trajectory.jsonl

litellm-trajectory.jsonl is the primary Tracer output. It is the replayable agent trajectory for one task and the source material for later conversion. result.json at the job root summarizes rewards, errors, and aggregate rollout status for dashboards and debugging.

SFT conversion outputs

When sft_conversion.enabled is true, Tracer converts successful rollouts through swe_data_process and writes:

artifacts/sft_data/<job>/
├── im.jsonl
├── lf.json
├── lf.stats.json
└── .convert_sig.json

lf.json is the training-ready LLaMA-Factory file. im.jsonl keeps the richer intermediate records and quality metadata used by dashboards and filtering.

Run history

Each completed run can be archived under:

artifacts/archives/run_NNN/
├── metadata.yaml
├── config.yaml
└── scripts/

artifacts/index.yaml records the run timeline. Use it for history; use the active job directory for live rollout details.

The launch script writes an archive entry after success, failure, or interruption, so interrupted jobs still leave a run snapshot for later inspection.