Research Dashboard

A real-time web interface for managing experiment campaigns, monitoring workers, and analysing results.

Overview

The XeeNet dashboard is the control centre for the platform. Researchers use it to create experiment briefs, monitor distributed training progress, and analyse results across campaigns. Built with HTMX for real-time updates, Jinja2 templates, and Pico CSS for a clean, responsive interface.

Dashboard overview showing workers, tasks, families, briefs, and total credits
Dashboard home: live platform stats showing 2 workers online, 68 total tasks, and 538 credits in circulation

Research Briefs

Research briefs are the starting point for every experiment campaign. A brief defines the research goal, compute budget, and target hardware profile. The orchestrator automatically decomposes each brief into concrete tasks with unique hyperparameter configurations.

Research briefs list showing experiment campaigns
Briefs list: each brief describes an experiment campaign with goal, budget, and hardware requirements

Campaign Results

The brief detail page shows the full lifecycle of a campaign: from the research goal through to aggregated results. Key metrics include best val_bpb, mean/standard deviation across all runs, and the winning hyperparameter configuration.

Brief detail page showing campaign results with best val_bpb, configuration, and task list
Campaign results: 10/10 tasks completed, best val_bpb 3.5705 with full hyperparameter config and per-task breakdown

Experiment Families

Each brief generates an experiment family (campaign) containing all related tasks. Families track the metric being optimised, completion status, and linked brief.

Experiment families list showing campaign status
Families view: campaigns linked to briefs with val_bpb metric tracking and completion status
Family detail showing linked brief, metric, resources, and task list
Family detail: linked brief, minimum resource requirements, and all 10 tasks with seed and time budget assignments

Task Management

Tasks are the atomic units of work. Each task represents a single training run with a specific hyperparameter configuration, time budget, and seed. The tasks view shows all tasks across all campaigns with their completion status.

Tasks list showing all training tasks with completion status
Tasks list: all training tasks across campaigns, showing type, family, priority, assigned worker, and completion time
Task detail showing hyperparameter config, resource requirements, and training result with val_bpb metric
Task detail: hyperparameter config, resource requirements, and real training result (val_bpb: 3.8885, train_loss: 1.7409)

Worker Monitoring

The workers page tracks all registered nodes in the network. Each worker reports its hardware profile (CPU cores, RAM, GPU model and VRAM), platform, and heartbeat status.

Workers list showing two online workers with hardware profiles
Connected workers: Intel Arc Pro 140T (16GB) and NVIDIA GeForce RTX 3080, both online on Windows

Technical Implementation

Template Architecture

The dashboard uses ~15 Jinja2 templates with a shared base layout. HTMX attributes handle dynamic content loading without full page refreshes:

Template Purpose
_base.html Base layout with navigation, Pico CSS, HTMX includes
home.html Platform overview with live statistics
briefs_list.html All research briefs with status badges
brief_detail.html Campaign results, best config, task breakdown
workers_list.html Connected workers and hardware profiles
credits_overview.html Credit balances and transaction history
_stats_partial.html HTMX partial for auto-refreshing statistics

No Build Step

The dashboard is pure server-side rendering. Static CSS and minimal JavaScript live in static/. HTMX handles dynamic updates via HTML-over-the-wire. This eliminates the entire frontend build pipeline: no webpack, no bundler, no node_modules. Run uvicorn services.api.main:app and the dashboard is ready.