Job list (/jobs)
Real-time view of all jobs: status, method, GPU count, duration, retry count, sweep badges. Filter by status / user / project. Inline actions: stop, pause, resume, delete.Jobs list. Filter by status or project; click a row for details.
Job details
Click into any job to see six tabs:Charts
Training scalar time series: loss, reward, KL, gradient norm, learning rate… Features:- Width-adaptive downsampling (LTTB), so large runs stay smooth;
- Step-window zooming and incremental refresh (new points appended every few seconds while running);
- Metric alias normalization (
train_loss/lossdisplayed as one source); - Pin charts to the first slots in their group; pin order is stored per project in this browser and reused across that project’s runs;
- CSV export.
env/tool_success_rate) are displayed alongside training scalars from the same source.
Charts tab: reward, accuracy, and train scalars while a job is running.
Logs
Full log stream: live SSE follow while running (auto-resume on disconnect), historical window playback, full-text search, download. A ring buffer protects against extremely long logs, so the browser never gets dragged down.Logs tab: live follow, search, download.
Validation
Per-round validation sample browsing: multi-turn conversation traces, reward distributions, comparison with the previous round. For RL methods, per-sample reward breakdowns are visible.Validation tab: per-round samples and reward breakdowns.
System
Hardware time series: GPU utilization, memory, temperature, network throughput. The first place to check whether “training is slow” is a data bottleneck or a communication bottleneck.Diagnostics
AI diagnostics: on failure, or on demand. It reads the log tail, metric trends, validation samples, and the config snapshot, then suggests knobs (including a resubmit plan).Lineage / Artifacts
Checkpoints / hf_exports / evaluation reports registered by this run, plus the upstream/downstream chain via--init-from.
Artifacts tab: checkpoints, exports, and lineage.