eval-monitoring

name	eval-monitoring
description	Guide for monitoring running SABER evaluations, checking progress, and managing eval batches. Use this when asked to monitor evals, check progress, produce a status report, or manage concurrent eval runs. Also covers Docker health and resource management.

Eval Monitoring Skill

Quick Status Check

Run the monitoring script for an instant report:

./scripts/monitor_evals.sh

For continuous monitoring (every 15 minutes):

./scripts/monitor_evals.sh --watch 15

Manual Monitoring Commands

1. Check Running Processes

# Count running evals
ps aux | grep "inspect eval" | grep -v grep | wc -l

# See model and agent for each
ps aux | grep "inspect eval" | grep -v grep | \
  grep -oP '(?<=--model )"?[^\s"]+' | tr -d '"'

2. Check Eval File Progress

Staleness note: .eval ZIP files are flushed in batches by inspect_ai, so sample counts read from them lag 2–3 minutes behind the live results visible in the inspect view UI. This is expected — treat .eval reads as "near-real-time" rather than exact.

# Quick sample count for an in-progress eval
python3 -c "
import zipfile, json
with zipfile.ZipFile('path/to/file.eval') as z:
    samples = [n for n in z.namelist() if n.startswith('samples/') and n.endswith('.json')]
    header = json.loads(z.read('header.json'))
    print(f'Status: {header.get(\"status\")}, Samples: {len(samples)}')
"

3. Check Docker Health

# Container count (>120 = resource pressure)
docker ps -q | wc -l

# Check for stuck containers
docker ps --format '{{.Names}} {{.Status}}' | grep -i "unhealthy\|restarting"

4. Check Log Files for Errors

# Find recently active logs
find latest_experiments/ -name "*_log.txt" -mmin -15 -exec tail -3 {} +

# Check a specific eval log
tail -20 latest_experiments/excytin/sonnet_4.6_claude_code_log.txt

Batch Running

Launch All Missing Evals (3 concurrent)

./scripts/run_eval_batch.sh --domain excytin --max-concurrent 3

Dry Run (see what would be launched)

./scripts/run_eval_batch.sh --domain excytin --dry-run

Override Concurrency

# Conservative: 2 at a time
./scripts/run_eval_batch.sh --max-concurrent 2

# Only if Docker resources allow: 4 at a time
./scripts/run_eval_batch.sh --max-concurrent 4

Concurrency Guidelines

Concurrent Evals	Docker Resources	When to Use
1	Minimal (~20 containers)	Debugging, resource-constrained
2-3	Moderate (~60-100)	Recommended default
4-5	Heavy (~100-160)	Only with fast models + good hardware
6+	Excessive (160+)	AVOID — Docker compose timeouts

Key rule: Never run more than 3 concurrent evals with --max-samples 64. Each eval can spawn up to 16 Docker sandbox containers simultaneously, plus the permanent services.

Eval File Naming Convention

Eval files follow this pattern: {model_slug}_{agent}.eval

Model Slug	Model ID
sonnet_4.6	anthropic/claude-sonnet-4-6
opus_4.6	anthropic/claude-opus-4-6
haiku_4.5	anthropic/claude-haiku-4-5
gpt_5.4	openai/azure/gpt-5.4
gpt_5.4_mini	openai/azure/gpt-5.4-mini

Agent	Description
react	Default React agent (step-by-step reasoning)
copilot	GH Copilot agent
claude_code	Claude Code agent

Interpreting Eval File Status

File Size	Status	Action
<50KB	Error/empty	Delete and retry
50KB-1MB	Partial (few samples)	Delete and retry
>1MB	Likely has data	Check header.json status field
Status: success	Complete	Skip
Status: error	Failed mid-run	Check error, may need retry
Status: started	In progress or abandoned	Check if process alive

Common Failure Modes

Docker Compose Timeout

Symptom: TimeoutError: Docker compose command ... timed out after 45 seconds Cause: Too many concurrent evals or Docker resource exhaustion Fix: Kill all evals, docker container prune -f, relaunch with lower concurrency

DeploymentNotFound (Azure)

Symptom: DeploymentNotFound in pre-flight check Cause: Wrong deployment name or API version Fix: Check deployment name matches Azure portal. Use AZUREAI_OPENAI_API_VERSION override.

Safety Refusals (0% scores)

Symptom: All scores 0.0, answers contain "I cannot", "social engineering" Cause: Model safety guardrails triggered by benchmark prompts Fix: Document in model_safety_filters_analysis.ipynb, not a retry-able error

Docker Reset Procedure

When Docker is in a bad state:

# 1. Gracefully stop all eval processes (allow cleanup of .eval files and Docker resources)
pkill -f "inspect eval"
sleep 10  # Give processes time to flush and clean up

# 2. Force-kill only if any are still running
pkill -9 -f "inspect eval" 2>/dev/null

# 3. Stop all containers (xargs -r skips if no containers are running)
docker ps -q | xargs -r docker stop

# 4. Remove stopped containers
docker container prune -f

# 5. Verify clean state
docker ps -q | wc -l  # Should be 0

# 6. Relaunch evals
./scripts/run_eval_batch.sh --domain excytin --max-concurrent 3