security-verifiers-utils 0.2.0

Shared utilities for Security Verifiers RL environments

security-verifiers-utils

Shared utilities and components for Security Verifiers RL environments.

Overview

This package provides common functionality used across all Security Verifiers environments, including:

• Dataset Loading: Multi-tiered dataset loading with automatic fallback (local → HuggingFace → synthetic)
• Response Parsing: Strict JSON schema validation for classification outputs
• Reward Functions: Reusable reward components for accuracy, calibration, and asymmetric costs
• Logging: Structured rollout logging with Weave and Weights & Biases integration
• Weave Integration: Automatic tracing initialization for Verifiers operations

Installation

pip install security-verifiers-utils
# or with uv
uv add security-verifiers-utils

Usage

Dataset Loading

from sv_shared import load_dataset_with_fallback, DatasetSource

# Automatic fallback: local → hub → synthetic
dataset = load_dataset_with_fallback(
    dataset_name="my-dataset.jsonl",
    local_path="./data",
    hf_repo="org/repo",
    dataset_source="auto",  # or "local", "hub", "synthetic"
    synthetic_factory=lambda: [{"example": "data"}],
    max_examples=1000,
)

Response Parsing

from sv_shared import JsonClassificationParser

parser = JsonClassificationParser(
    allowed_labels=["Benign", "Malicious", "Abstain"]
)

# Parse and validate JSON response
result = parser(response)
# Returns: {"label": str, "confidence": float, "rationale": str}
# Or None if invalid

Reward Functions

from sv_shared import (
    reward_accuracy,
    reward_calibration,
    reward_asymmetric_cost,
)

# Accuracy reward (0.0 to 1.0)
acc_reward = reward_accuracy(
    predicted="Malicious",
    ground_truth="Malicious",
    abstain_label="Abstain"
)

# Calibration reward (encourages well-calibrated confidence)
cal_reward = reward_calibration(
    predicted="Malicious",
    ground_truth="Malicious",
    confidence=0.85,
    abstain_label="Abstain"
)

# Asymmetric cost (penalizes false negatives more than false positives)
cost_reward = reward_asymmetric_cost(
    predicted="Benign",
    ground_truth="Malicious",
    fn_cost=10.0,  # False negative cost
    fp_cost=1.0,   # False positive cost
    abstain_label="Abstain"
)

Rollout Logging

from sv_shared import build_rollout_logger

# Build logger with Weave and W&B backends
logger = build_rollout_logger({
    "weave_project": "my-project",
    "wandb_project": "my-project",
    "wandb_entity": "my-org",
})

# Use with environment
env = load_environment(logger=logger)

Weave Auto-tracing

# Import before verifiers to enable automatic tracing
from sv_shared import weave_init  # Initializes Weave if enabled
import verifiers as vf

# Configure via environment variables:
# WEAVE_AUTO_INIT=true/false (default: true)
# WEAVE_PROJECT=<name> (default: security-verifiers)
# WEAVE_DISABLED=true/false

Components

Dataset Loader (dataset_loader.py)

• load_dataset_with_fallback() - Multi-tiered dataset loading
• DatasetSource - Type alias for dataset source modes
• DEFAULT_E1_HF_REPO, DEFAULT_E2_HF_REPO - Default HuggingFace repositories

Parsers (parsers.py)

• JsonClassificationParser - Validates JSON classification outputs with strict schema adherence

Rewards (rewards.py)

• reward_accuracy() - Classification accuracy reward
• reward_calibration() - Confidence calibration reward
• reward_asymmetric_cost() - Asymmetric false positive/negative costs

Rollout Logging (rollout_logging.py)

• RolloutLogger - Base logger class
• build_rollout_logger() - Factory for creating loggers with backends
• RolloutLoggingConfig - Configuration dataclass
• DEFAULT_ROLLOUT_LOGGING_CONFIG - Default configuration

Weave Initialization (weave_init.py)

• initialize_weave_if_enabled() - Conditional Weave initialization
• Automatic tracing setup for Verifiers

Utilities (utils.py)

• get_response_text() - Extract text from Verifiers response objects

Development

Setup

git clone https://github.com/intertwine/security-verifiers.git
cd security-verifiers/sv_shared
uv venv && source .venv/bin/activate
uv sync --extra dev

Testing

pytest

Linting

ruff check .
ruff format .

Environment Variables

• WEAVE_AUTO_INIT - Enable/disable automatic Weave initialization (default: true)
• WEAVE_PROJECT - Weave project name (default: security-verifiers)
• WEAVE_DISABLED - Completely disable Weave (default: false)
• WANDB_API_KEY - Weights & Biases API key (required for W&B logging)
• HF_TOKEN - HuggingFace token (required for private dataset access)

License

MIT License - see LICENSE file for details.

Links

• Repository: https://github.com/intertwine/security-verifiers
• Issues: https://github.com/intertwine/security-verifiers/issues
• Documentation: https://github.com/intertwine/security-verifiers/blob/main/README.md

Related Packages

• verifiers - Core Verifiers RL framework
• sv-env-network-logs - Network log anomaly detection environment
• sv-env-config-verification - Configuration security verification environment