rldash

A zero-dependency terminal dashboard for RL training runs. One Python file, pure stdlib, works on Linux / macOS / Windows. Point it at your training log and it redraws in place: progress bars, return sparkline, steps/sec, and a GPU gauge if nvidia-smi is on your PATH.

╔════════════════════════════════════════════════════════════════╗
║   N = 7   P E N D U L U M                   r l d a s h         ║
║   n7_bal_long.log                                               ║
╚════════════════════════════════════════════════════════════════╝

  update      195/1907   ▕███▏░░░░░░░░░░░░░░░░░░░░░░░░░░▏  10.2%
  steps        102.2M / 1,000M      SPS 327,142

  ep_ret   ▕██████████████████████████████▏     217.1  (peak 217)
  rew/step ▕████████████████████▏░░░░░░░░░▏      2.35
  ep_len   ▕████████████████████████▏░░░░░▏        92

  ep_ret trend  ▁▂▃▄▅▆▇█▇▆▇█  (last 48)

  ┌─ GPU ───────────────────────────────────────────────┐
  │ temp   64°C ▕███████████████▏░░░░▏ limit 85         │
  │ power   304 W   util  48%                           │
  └─────────────────────────────────────────────────────┘
  elapsed 00:42:10  ·  11:24:28  ·  training  ·  Ctrl-C to quit

Quickstart

Grab the one file and run it:

curl -O https://raw.githubusercontent.com/bmdhodl/rldash/main/rldash.py
python rldash.py --log "runs/*.log"

No install, no dependencies. With a glob, it always follows the most recently modified match — kick off a new run and the dashboard hops over to it automatically.

Your log format

Out of the box it parses lines like:

upd  205/1907  step 107,479,040  SPS 327,723  ep_ret 220.3  ep_len 94.8  rps 2.32 ...

Different trainer? Pass your own regex with named groups. Only step is required; upd, updtot, sps, ep_ret, ep_len, and rps light up gauges when present — and any other named group you add shows up automatically in the diagnostics footer (v_loss, EV, logstd, whatever your trainer prints):

python rldash.py --log train.log \
  --pattern "global step (?P<step>\d+).*?mean reward (?P<ep_ret>-?[\d.]+)"

Useful flags:

flag	what it does
--title "MY RUN"	header text
--interval 1.0	refresh seconds
--ep-len-max 650	fixed full-scale for the ep_len bar
--rps-max 6	fixed full-scale for the reward/step bar
--gpu-limit 80	temperature gauge limit
--done-pattern "COMPLETE|Traceback"	run-state markers: COMPLETE vs CRASHED in the footer
--plain --once	print one frame and exit (CI, piping)

Training inside WSL, watching from Windows?

Two good options:

# point Windows Python at the WSL filesystem
python rldash.py --log "\\wsl$\Ubuntu-22.04\home\<you>\myproj\runs\*.log"

or just run rldash inside WSL next to the run — it's pure stdlib, it runs anywhere. (If no log matches, rldash prints exactly where it looked and these same hints.)

PowerShell flavor (Windows + WSL)

The original lives in powershell/: a PowerShell dashboard that watches training running inside WSL from a native Windows console, via a small WSL-side snapshot script. Same gauges, same vibe:

powershell -NoExit -ExecutionPolicy Bypass -File powershell\watch_rl.ps1

Where this came from

Built live during an N=7 (seven-link) pendulum swing-up campaign on an RTX 5090 — billions of PPO steps that needed watching without babysitting raw logs. The details (GPU temp limit front and center, sparkline of returns, auto-following the newest run) are exactly the things we kept wanting mid-campaign.

More builds and write-ups: bmdpat.com

License

MIT