brycewang-stanford/training-check

Training Check

Periodically read WandB metrics during training to catch problems early. Do not wait until training finishes to discover it was a waste of GPU time.

Context: $ARGUMENTS

Constants

• WANDB_RUN - Read from project notes or pass as entity/project/run_id.
• CHECK_INTERVAL - Starts at 10 minutes, then gradually increases if consistently healthy: 10 min -> 20 min -> 30 min -> 60 min (cap).
• REVIEWER_MODEL = gpt-5.4 - Used via a secondary Codex agent for ambiguous cases only.

When to Use

• After training is confirmed running (session alive, loss decreasing for the first few steps)
• When the user wants recurring health checks during training
• This skill checks training QUALITY, not process HEALTH. Process health (session alive, GPU utilization) belongs to watchdog-style monitoring.

Workflow

Step 1: Read WandB Metrics

import wandb
api = wandb.Api()
run = api.run("<entity>/<project>/<run_id>")
history = run.history()

If WandB is unreachable (API error, network issue), fall back to reading the log file directly via SSH:

ssh server "tail -100 /path/to/training.log"

Check these signals:

• Loss trend - Is training loss decreasing over the last N steps?
• Eval metrics - Are evaluation metrics improving (or at least not degrading)?
• NaN / Inf - Any NaN or Inf values in loss or gradients?
• Spikes - Sudden large jumps in loss (>10x normal variance)?
• Learning rate - Is the schedule behaving as expected?
• Gradient norm - Exploding or vanishing?

Step 2: Judgment

| Signal | Judgment | Action | |--------|----------|--------| | NaN/Inf in loss | Clearly bad | Stop training, investigate | | Loss diverging (increasing for >N steps) | Clearly bad | Stop training, investigate | | Eval metrics significantly worse than baseline | Clearly bad | Stop training, investigate | | Loss decreasing, metrics improving | Clearly fine | Continue, increase check interval | | Loss flat but not diverging | Unsure | -> Step 3 (secondary review) | | Metrics noisy, can't tell trend | Unsure | -> Step 3 (secondary review) | | Slightly worse than baseline but still early | Unsure | -> Step 3 (secondary review) |

Step 3: Secondary Codex Judgment (only when unsure)

Only escalate when the signal is ambiguous. For clearly good or clearly bad signals, act directly.

spawn_agent:
  model: REVIEWER_MODEL
  reasoning_effort: high
  message: |
    TRAINING HEALTH CHECK - need your judgment on ambiguous metrics.

    Run: <entity>/<project>/<run_id>
    Current epoch/step: X / Y total
    Training loss (last 10 checkpoints): [values]
    Eval metrics (last 3 evals): [values]
    Baseline reference: [numbers from paper/reproduction]

    What I'm unsure about: [specific concern]

    Please respond with exactly one of:
    - STOP: clearly problematic, should kill training
    - CONTINUE: looks fine, check again next interval
    - WAIT: not enough data to judge, check again sooner

If delegation is unavailable, make a local judgment using the same rubric and mark the decision [pending external review]. In ambiguous cases with no hard failure, prefer WAIT over STOP.

Step 4: Act

| Decision | Action | |----------|--------| | Stop | Kill the training session. Save the WandB run URL, key metrics, and reason for stopping. Log to project notes for debugging. | | Continue | Do nothing. Re-run at the next interval (increase interval if consistently healthy). | | Wait | Do nothing but keep the current short interval (do not increase). |

Integration with Watchdog

training-check and watchdog-style monitoring operate at different levels:

| Layer | Tool | What it checks | Frequency | |-------|------|----------------|-----------| | Process health | watchdog | Session alive? GPU active? | Every 60s (continuous) | | Training quality | training-check | Loss trend? Metrics improving? | Every 10-60 min (periodic) |

Use both together:

• Watchdog catches crashes and idle GPUs immediately
• training-check catches subtle quality issues (loss plateau, metric degradation)

Rules

• Do not stop training on the first sign of noise - some loss spikes are normal. Look at trends over multiple checkpoints.
• When stopping training, always save the WandB run URL and key metrics as evidence.
• If both WandB and log files are unreachable, report the connectivity issue and try again next interval. Do not assume training is broken.
• Gradually increase check interval when healthy (10 -> 20 -> 30 -> 60 min). Reset to 10 min after any anomaly.
• This skill is meant to be automated via a recurring scheduler. If the user wants ongoing monitoring, set up the best local mechanism available instead of waiting for manual reruns.

Recurring Setup Example

After training is confirmed stable:
  Create a recurring job (cron, task scheduler, tmux loop, etc.)
  that runs `/training-check <entity>/<project>/<run_id>` every 10 minutes.

As the check interval increases, update the old recurring job to match the new interval.