Skip to content

README.md

Latest commit

 

History

History
90 lines (78 loc) · 3.11 KB

README.md

File metadata and controls

90 lines (78 loc) · 3.11 KB
Raw

Evaluation Scripts

This directory contains evaluation scripts for testing LLM guards and evaluators on mutated trace datasets.

Running Evaluations

1. evaluate_guard.py (Evaluating specialized guard models)

Open Source Models (e.g., Llama-Guard 3): First, host the model via vllm:

CUDA_VISIBLE_DEVICES=0 vllm Qwen3/Llama-Guard-3-8B \
    --served-model-name llama_guard \
    --port 8001 \
    --gpu-memory-utilization 0.45 \
    --max-model-len 12288 \
    --dtype auto \
    --api-key "EMPTY"

Run the evaluation script connecting to it:

python evaluate_guard.py --model_name llama_guard --model_family llama --base_url http://localhost:8001/v1 --output_dir ./results/Llama3-8B-Guard

Azure Content Safety:

python evaluate_guard.py --provider azure \
    --azure_endpoint "https://foundry-ml.cognitiveservices.azure.com/" \
    --azure_key "your_key" \
    --output_dir ./results/Azure-Guard

GCP Model Armor:

python evaluate_guard.py \
    --provider gcp \
    --gcp_project <project> \
    --gcp_region <region> \
    --gcp_model_name <model_name> \
    --gcp_key_file <file_name>.json \
    --output_dir ./results/GCP-Guard \
    --num_samples 10

2. evaluate_llm.py (Evaluating versatile LLMs)

Open Source LLM Evaluator (e.g., Qwen 32B): Host via vllm:

CUDA_VISIBLE_DEVICES=2 vllm serve Qwen3/Qwen3-32B \
    --port 8017 \
    --gpu-memory-utilization 0.9 \
    --max-model-len 12288 \
    --dtype auto \
    --api-key "EMPTY"

Evaluate:

python evaluate_llm.py --model_name "Qwen3/Qwen3-32B" --api_key "EMPTY" \
    --base_url "<base_url>" \
    --output_dir "./results/Qwen3-32B"

OpenAI/Proprietary LLM Evaluator:

python evaluate_llm.py --model_name gpt-5.4-mini \
    --extra_params '{"reasoning": {"effort": "low"}}' \
    --base_url "<base_url>" \
    --output_dir "./results/GPT5.4-mini-medium"

Script Arguments Reference

evaluate_llm.py

  • --model_name: Name of the LLM to evaluate.
  • --api_key: API key for the model (uses OPENAI_API_KEY by default).
  • --base_url: Base URL for the API (uses OPENAI_BASE_URL by default).
  • --settings: Assessment modes (e.g., fine_grained_classification, coarse_grained_classification).
  • --output_dir: Base directory for storing benchmark summaries.
  • --num_samples: Cap on entries to evaluate per category (0 for all).
  • --workers: Max parallel workers utilizing core_utils.runner.
  • --extra_params: JSON string for explicit API-kwargs (e.g., {"reasoning_effort": "low"}).

evaluate_guard.py

  • --provider: Specifies the SDK adapter (openai_compatible, azure, aws, gcp).
  • --model_name: Fallback model name for specific APIs.
  • --model_family: For openai_compatible guards to dictate safety-check parsing (llama, qwen, granite).
  • --azure_endpoint / --azure_key / --azure_severity_threshold: Configurations for Azure Content Safety.
  • --aws_region / --aws_guardrail_id / --aws_guardrail_version: Configurations for AWS Bedrock.
  • --gcp_project / --gcp_region / --gcp_model_name / --gcp_key_file: Configs for GCP Model Armor.