BlocksTrainer
Dashboard
Trainer's dashboard is a read-only monitor for LLaMA-Factory training runs. Run
/trainer:dashboard to check training health, compare runs, and inspect final
metrics. It never launches, pauses, resumes or edits a job; use /trainer:run
for training.
A demo dashboard is available here.
Deployment mode
Both modes read the same files, so the numbers never depend on which one you pick.
- Local (default): run
/trainer:dashboard. You get a status summary in the terminal, and the web UI on request. Everything stays on the training host, which matters because logs and config snapshots carry model paths, dataset names and host paths. - Cloud: run
/trainer:dashboard publish to Cloudflare. SetCLOUDFLARE_API_TOKEN,CLOUDFLARE_ACCOUNT_IDand the target Pages project first. For anything long-lived, prefer an access-controlled setup, since the URL alone exposes the run.
Where it reads from
The board reads two fixed locations, both written by scripts/train.sh:
| Path | What it holds |
|---|---|
artifacts/model/ | Training runs, one per immediate subdirectory, including checkpoints. Progress comes from each run's trainer_log.jsonl and trainer_state.json, and the final numbers from its *_results.json. |
artifacts/logs/ | Console logs from the launcher, which is where to look when a run dies before writing any metrics. |
Import External Data
The board only reads artifacts/. Symlink an outside run into
artifacts/model/, and the link name becomes the run name.
What it monitors
| Monitor | What it tells you |
|---|---|
| Run progress | Current step, epoch, percentage, ETA, and whether the run is active or finished. |
| Training loss | Loss curve, for spotting divergence, instability, or stalled learning. |
| Learning rate | Scheduler behavior, warmup, and decay across steps. |
| Gradient norm | Stability signal for catching exploding gradients. |
| Performance | Step time, samples/sec, runtime, throughput, and FLOPs. |
| Evaluation metrics | Eval loss and best checkpoint, when the run includes an eval split. |
| Run comparison | Side-by-side comparison of runs, datasets, models, or hyperparameters. |
| Logs and artifacts | Console logs, result JSONs, trainer state, checkpoints, and loss plots. |
When to Use It
Use /trainer:dashboard when you want to answer questions like:
- Is the training run still alive?
- Is loss decreasing normally?
- Did throughput or step time suddenly change?
- Which checkpoint is the latest or the best?
- How does this run compare with previous data or model settings?
For training mechanics, see Design. For artifact contracts, see Output Format.