Dashboards
CLaaS includes two built-in web dashboards served by the API. No separate frontend is needed. They are rendered as HTML responses.
Training Dashboard
URL: http://localhost:8080/v1/dashboard
The training dashboard shows recent feedback batches with a paginated table view.
Features
- Each row is a feedback batch, expandable to reveal per-sample details
- Per-sample view shows prompt, response, and feedback text
- Batch-level timing breakdown (total time, training time, inference time)
- SDPO training metrics per batch:
- Local mode: loss, KL regularization, importance sampling ratio, clip fraction
- Tinker mode:
adv_mean, kl_mean, and other Tinker-specific metrics
- Pagination:
?page=1&per_page=20 (max 200 per page)
Data source
The dashboard reads JSON log files from the FEEDBACK_LOG_DIR directory (default ./data/feedback). Each feedback submission creates a structured log entry.
Eval Dashboard
URL: http://localhost:8080/v1/eval?results_dir=./data/evals
The eval dashboard displays results from running the eval harness against your model.
Features
- Discovers all eval runs under the results directory (newest first)
- Each run is a collapsible section showing:
- Summary table: preference, LoRA ID, margin delta, compliance, capability ratio
- Embedded plots: base64-encoded matplotlib PNGs from the
plots/ subdirectory
- Per-preference step tables with expandable detail rows (step, timestamp, margin, delta, compliance, loss, timing)
- Expandable rollout transcripts for each step
- Baseline metrics (collapsible)
- Pagination:
?page=1&per_page=20 (max 200 per page)
Query parameters
| Parameter | Default | Description |
|---|
results_dir | ./data/evals | Path to the eval results directory |
page | 1 | Page number (1-indexed) |
per_page | 20 | Number of runs per page (max 200) |
The results_dir parameter must be within ./data/evals. Path traversal outside this directory is blocked.
API endpoints
Both dashboards are served as HTMLResponse from FastAPI:
GET /v1/dashboard → Training dashboard
GET /v1/eval → Eval dashboard
