Security Verifiers Evaluation

Run reproducible evaluations on E1 (network-logs) and E2 (config-verification) environments, generate metrics reports, and analyze results.

Prerequisites

Ensure environment variables are set (check .env file):

• OPENAIAPIKEY - For OpenAI models (gpt-*)
• OPENROUTERAPIKEY - For non-OpenAI models via OpenRouter
• WANDBAPIKEY - For Weave logging (optional)

Running Evaluations

E1: Network Logs (Classification)

Single-turn anomaly detection with calibration and asymmetric costs.

# Basic eval (10 examples)
make eval-e1 MODELS="gpt-5-mini" N=10

# Multiple models
make eval-e1 MODELS="gpt-5-mini,gpt-4.1-mini,qwen3-14b" N=100

# Full production dataset
make eval-e1 MODELS="gpt-5-mini" N=1800 DATASET="iot23-train-dev-test-v1.jsonl"

# OOD datasets
make eval-e1 MODELS="gpt-5-mini" N=600 DATASET="cic-ids-2017-ood-v1.jsonl"
make eval-e1 MODELS="gpt-5-mini" N=600 DATASET="unsw-nb15-ood-v1.jsonl"

E2: Config Verification (Tool-Use)

Multi-turn tool-grounded auditing with KubeLinter, Semgrep, and OPA.

# Basic eval with tools (2 examples)
make eval-e2 MODELS="gpt-5-mini" N=2 INCLUDE_TOOLS=true

# Without tools (single-turn)
make eval-e2 MODELS="gpt-5-mini" N=10 INCLUDE_TOOLS=false

# Dataset options
make eval-e2 MODELS="gpt-5-mini" N=50 DATASET="k8s-labeled-v1.jsonl"
make eval-e2 MODELS="gpt-5-mini" N=50 DATASET="terraform-labeled-v1.jsonl"
make eval-e2 MODELS="gpt-5-mini" N=100 DATASET="combined"  # default

Error Handling

Early stopping prevents wasted API costs:

# Stop after 5 consecutive errors (default: 3)
make eval-e1 MODELS="gpt-5-mini" N=100 MAX_CONSECUTIVE_ERRORS=5

# Disable early stopping
make eval-e1 MODELS="gpt-5-mini" N=100 MAX_CONSECUTIVE_ERRORS=0

Generating Reports

SV-Bench supports two kinds of reporting:

• Per-run contract reports: summary.json (schema-stable) + report.md (human readable)
• Cross-run comparison reports: timestamped report-*.json aggregations for quick comparisons

Per-run SV-Bench report (contract-grade)

# E1
WEAVE_DISABLED=true .venv/bin/svbench_report --env e1 --input outputs/evals/sv-env-network-logs--gpt-5-mini/<run_id> --strict

# E2
WEAVE_DISABLED=true .venv/bin/svbench_report --env e2 --input outputs/evals/sv-env-config-verification--gpt-5-mini/<run_id> --strict

E1 comparison report (legacy-friendly JSON table)

# All non-archived runs
make report-network-logs

# Specific runs
make report-network-logs RUN_IDS="run_abc123 run_def456"

# Custom output path
make report-network-logs OUTPUT="reports/e1-comparison.json"

E2 comparison report (legacy-friendly JSON table)

make report-config-verification
make report-config-verification RUN_IDS="run_abc123"

Understanding Results

Output Structure

outputs/evals/sv-env-{name}--{model}/{run_id}/
├── metadata.json    # Run config, versions, git SHA
├── results.jsonl    # Per-example results
├── summary.json     # SV-Bench schema summary (contract-grade)
└── report.md        # Human-readable report

Note: OpenRouter models create nested directories:

outputs/evals/sv-env-network-logs--qwen/qwen-2.5-7b-instruct/{run_id}/
outputs/evals/sv-env-network-logs--meta-llama/llama-3.1-8b-instruct/{run_id}/

The report scripts use recursive glob to find all runs regardless of nesting depth.

Key E1 Metrics (network-logs)

Key E2 Metrics (config-verification)

Model Routing

Model names are auto-resolved via scripts/model_router.py:

• OpenAI models: gpt-5-mini, gpt-4.1-mini, o1-mini
• OpenRouter models: qwen3-14b → qwen/qwen3-14b, llama-3.1-8b → meta-llama/llama-3.1-8b-instruct

Comparing Models

1. Run evals with same parameters:

make eval-e1 MODELS="gpt-5-mini,gpt-4.1-mini,qwen3-14b" N=500

1. Generate report:

make report-network-logs

1. Review summary.json in each run directory for per-model metrics.

Troubleshooting

Rate limits: Reduce N or use MAXCONSECUTIVEERRORS. Missing API key: Check .env has correct key for model provider. Model not found: Use full OpenRouter path (e.g., openai/gpt-5-mini).

FormatSuccess% = 0%

If a model shows 0% format success, common causes:

1. Markdown code blocks: Many models (gpt-4o-mini, qwen, llama) wrap JSON in `json...` blocks. The parsers handle this automatically via extractjsonfrommarkdown() in svshared/parsers.py.

1. Model outputs prose instead of JSON: Some models explain what they would do rather than returning the required JSON format. This is model behavior, not a parsing bug. Check results.jsonl completions.

1. Empty prompts: If prompt field in results.jsonl is empty, the dataset loading may have failed. Check that:

- Hub datasets have correct schema (question field) - Local JSONL files exist in environments/sv-env-*/data/ - HF_TOKEN is set for Hub access

Investigating Failed Runs

# Check a specific result file
cat outputs/evals/sv-env-*--{model}/{run_id}/results.jsonl | head -1 | python -m json.tool

# Look for empty prompts (indicates dataset loading issue)
grep '"prompt": ""' outputs/evals/sv-env-*--{model}/{run_id}/results.jsonl

# Check completion format
cat outputs/evals/sv-env-*--{model}/{run_id}/results.jsonl | python -c "
import json, sys
for line in sys.stdin:
    r = json.loads(line)
    print(f'Index {r[\"index\"]}: format_reward={r[\"rewards\"][\"format_reward\"]}')
    print(f'  completion: {repr(r[\"completion\"][:100])}...')
"