A Lightweight Validation Framework for Trustworthiness Assessment of Open-Source Large Language Models

This article introduces an open-source LLM trustworthiness assessment framework developed by merma1509. It centers on three core dimensions—security and refusal mechanisms, authenticity and hallucination detection, consistency and robustness—to build a lightweight and reproducible evaluation system. Based on the Ollama platform, this framework supports local deployment without the need for expensive API calls or high-end GPUs, lowering the validation threshold for small and medium teams as well as individual developers, and providing a practical tool for model trustworthiness assessment.

With the widespread deployment of LLMs in practical applications, trustworthiness (including security, authenticity, and consistency) has become a key concern. However, existing mainstream evaluation benchmarks mostly focus on capability performance and lack systematic assessment of trustworthiness indicators such as security refusal and hallucination generation. Additionally, most evaluation schemes rely on expensive APIs or high-performance GPUs, making it difficult for small and medium teams to conduct independent validation. Thus, there is an urgent need for lightweight and low-cost trustworthiness assessment tools.

The framework is structured around three core dimensions:

1. Security and Refusal Mechanisms: Test the model's ability to resist malicious prompts (e.g., injection attacks, role hijacking) by comparing response differences between 20 adversarial prompts and 10 benign prompts;
2. Authenticity and Hallucination Detection: Examine whether the model openly expresses its boundaries when facing uncertain information, using 20 adversarial prompts and 10 benign prompts;
3. Consistency and Robustness: Evaluate response stability under repeated queries or input perturbations, including 15 sets of repeated prompts and 5 benign controls. The framework uses pure local inference (based on Ollama) without external APIs. It generates a comprehensive score through weighted aggregation and introduces confidence interval calculation and weight sensitivity analysis to ensure robust results.

The technical architecture is modularly designed, with core components including the Ollama API client, dimension-specific evaluation modules, and result aggregation modules. The code is based on Python 3.11+ and depends on libraries such as requests and numpy. The data layer uses JSONL to store test prompts and responses, supporting custom use case expansion. The evaluation process can be run with a single command via run_evaluation.py, outputting dimension scores, 95% confidence intervals, weight sensitivity analysis, and original response archives. Methodologically, fixed random seeds ensure reproducibility, the Bootstrap method is used to calculate confidence intervals, and weight sensitivity analysis tests five configurations to avoid conclusion bias.

This framework has practical value for open-source LLM maintainers and users:

• Model developers can use it for pre-release self-inspection to identify security vulnerabilities and hallucination tendencies;
• Enterprise users can conduct independent trustworthiness tests during the selection phase to supplement risks not covered by official benchmarks;
• Its lightweight nature makes it suitable for integration into CI/CD pipelines, enabling continuous trustworthiness monitoring and rapid detection of regression issues in version updates.

The current version has limitations: it focuses on three core dimensions with a streamlined test scale; it mainly targets English scenarios with limited multilingual support. Future expansion directions include: adding evaluation dimensions such as fairness and privacy protection; expanding the test dataset scale; supporting more model backends; and improving cross-language trustworthiness assessment capabilities.