garak vs SWE-bench
garak
garak.aiSWE-bench
swebench.comSide by side
| Figure | garak | SWE-bench |
|---|---|---|
| Price from | Free | — |
| Free plan | Yes | — |
| Free trial | — | — |
| Open source | — | — |
| Web | — | Yes |
| Windows | Yes | Yes |
| macOS | Yes | Yes |
| Linux | Yes | Yes |
| API | — | Yes |
| Self-hosted | Yes | Yes |
| Plans published | 1 | 0 |
| Spec score · LLM Evaluation Tools | 7.8 (#12) | 7.0 (#25) |
| Integrations | The project lists support for OpenAI, Hugging Face, Cohere, Replicate, AWS Bedrock, LiteLLM, REST-accessible models, and custom Python integrations. | The project documents using Hugging Face datasets to load SWE-bench data and Modal for cloud evaluations. |
| Purpose | garak is an LLM vulnerability scanner that probes language models and systems using them for security weaknesses. | SWE-bench evaluates language models on real-world GitHub software issues by asking them to generate patches that resolve the issues. |
| Support | The project points users to its user guide, reference documentation, Discord, GitHub issues, and an NVIDIA email address for support. | The contact page directs general inquiries to the SWE-bench team by email. |
Plans
In their own words
garak
garak is a command-line scanner for probing language models and systems that use them for security weaknesses. Its checks cover issues such as hallucination, data leakage, prompt injection, misinformation, toxicity generation, and jailbreaks. Probes try to elicit failures, while detectors inspect responses for patterns such as unsafe content or a failure to refuse. A standard scan can run without supervision and produce a report, with scan settings available for customization. Integrations include OpenAI, Hugging Face, Cohere, Replicate, AWS Bedrock, LiteLLM, REST-accessible models, and custom Python integrations. Hugging Face models can run locally or be accessed through an inference API or private endpoint. Output includes screen results, a detailed report log, a hit log, and a debug log. garak assesses text output only. It is developed in Linux and OSX and can be installed with pip or from source; listed platforms include Linux, macOS, Windows, and self-hosted use. It is free, with a free plan and no free trial. Listed capabilities include custom metrics, LLM-as-a-judge, and safety evaluations.
SWE-bench
SWE-bench evaluates language models on software issues collected from GitHub. A model generates a patch, which the evaluation harness applies in a Docker environment before running the repository tests to check whether the issue is resolved. The original benchmark contains 2,294 task instances from 12 Python repositories. Other listed datasets include SWE-bench Lite, Verified, Multimodal and Multilingual. Verified contains 500 instances engineers checked as solvable; Multilingual has 300 instances across nine languages and 42 repositories. Multimodal includes 100 development instances with screenshots and interface elements, while test evaluation uses the SWE-bench API. Its CLI can submit predictions, retrieve reports, manage evaluation runs and track performance in the cloud. Evaluations can run locally or in the cloud using Modal and sb-cli. Installation requires Python 3.9 or later and Docker. Guides cover macOS, Linux and Windows, with WSL 2 required on Windows. The Docker guide recommends at least 120 GB of free disk space and 16 GB of RAM; caching all Docker images can require about 2,000 GB. SWE-bench is free.