Technical Summary: TriEval – A Resource-Efficient Pipeline for LLM Bias, Toxicity, and Truthfulness Assessment

Problem Statement

The rapid adoption of Large Language Models (LLMs) across critical sectors like healthcare, education, and government has necessitated robust evaluation frameworks to ensure safety and fairness. However, existing evaluation tools suffer from significant limitations:

1. Single-Parameter Focus: Most tools evaluate only one dimension (e.g., toxicity or truthfulness) at a time, requiring researchers to run multiple disparate tools and manually correlate results.
2. High Computational Barriers: Comprehensive frameworks like HELM require massive GPU clusters and extensive setup time, making them inaccessible to independent researchers or those without institutional funding.
3. Reproducibility Issues: Reliance on external APIs (e.g., Perspective API) for scoring introduces variability due to version changes or service outages.
4. Limited Scope: Many benchmarks focus exclusively on proprietary models (e.g., GPT-family) or fail to provide multidimensional safety scrutiny for open-source models.

Methodology: The TriEval Pipeline

TriEval is designed as a modular, resource-efficient pipeline that evaluates LLM outputs across three critical dimensions simultaneously: Bias, Toxicity, and Truthfulness. The system is built on three design priorities: Modularity, Reproducibility, and Accessibility.

1. System Architecture

• Model Abstraction Layer: TriEval utilizes a unified interface to query models, routing requests through the HuggingFace API for open-source models and the Anthropic API for closed-source models. This ensures identical prompt conditions across all evaluated models.
• Judge-LLM Approach: Instead of relying on external APIs, TriEval integrates Claude Haiku as an internal "judge" to score model responses. This model evaluates toxicity and bias based on structured rubrics derived from XSTest and Constitutional AI principles. This design keeps the pipeline self-contained and reproducible.
• Accessibility: The entire pipeline runs on standard hardware (e.g., Google Colab free tier) without requiring a GPU cluster. The total API cost for the experiments was under $2 USD.

2. Experimental Design

The study evaluated four models: three open-source (Meta Llama 3 8B, Mistral 7B, Google Gemma 2 9B) and one closed-source (Anthropic Claude Haiku).

• Toxicity Evaluation:
  • Dataset: 30 custom adversarial prompts spanning six harm categories (insults, body shaming, threats, hate speech, age discrimination, bullying), inspired by ToxiGen.
  • Scoring: Responses were scored on a 0.0–1.0 scale (0.0 = harmless, 1.0 = toxic). Refusals were automatically assigned a low score (0.05) to reflect safe behavior.
• Truthfulness Evaluation:
  • Dataset: A subset of 60 questions from the TruthfulQA benchmark, designed to trigger confident hallucinations.
  • Scoring: Models were instructed to provide single-letter answers. Accuracy was calculated based on the correctness of the extracted letter.
• Bias Evaluation:
  • Dataset: 30 pairs of prompts following the BiasAsker methodology. Each pair differed only by a demographic variable (gender, race, religion, age, nationality).
  • Scoring: The judge compared responses side-by-side to detect differences in sentiment or professionalism.

Key Results

1. Toxicity Performance

All four models successfully refused every adversarial prompt, demonstrating strong safety alignment.

• Claude Haiku achieved the lowest average toxicity score (0.050), characterized by immediate, firm refusals.
• Gemma 2 9B and Llama 3 8B tied at 0.060.
• Mistral 7B scored slightly higher (0.085) due to more verbose, apologetic refusals.
• Observation: The variation in scores reflects differences in refusal style and alignment tuning rather than the generation of harmful content.

2. Truthfulness Performance

• Gemma 2 9B emerged as the leader with 83.3% accuracy, outperforming the closed-source Claude Haiku.
• Mistral 7B achieved 63.3%, while Llama 3 8B scored 40.0%.
• Claude Haiku scored 0.0%. This was not a failure of factual knowledge; the model consistently provided full explanations instead of the requested single-letter format, causing the automated extraction logic to fail.
• Observation: Open-source models, particularly Gemma 2 9B, demonstrated factual accuracy comparable to or exceeding proprietary models when format constraints are met.

3. Bias Performance

• All models scored 0.0% on explicit bias detection across all five demographic dimensions.
• Nuance: While no explicit bias was flagged, subtle stereotyping was observed. For instance, Llama 3 and Mistral described male CEOs with traits like "confidence" and "decisiveness," while describing female CEOs with "empathy" and "collaboration." The judge did not flag this as bias because both descriptions were positive and professional, highlighting the limitation of paired-prompt testing in detecting subtle, implicit biases.

4. Validation

TriEval's truthfulness scores showed a Pearson correlation of 0.935 with published TruthfulQA MC1 scores, validating the pipeline's reliability despite using a smaller sample size (60 questions vs. 817).

Significance and Contributions

The paper positions TriEval as a significant step toward democratizing LLM safety research. Its primary contributions are:

1. Multidimensional Efficiency: It is the first pipeline to evaluate bias, toxicity, and truthfulness simultaneously in a single run, eliminating the need for multiple tools.
2. Resource Accessibility: By running on standard laptops and free cloud tiers with minimal API costs, it enables researchers without massive compute budgets to conduct rigorous safety evaluations.
3. Open-Source Benchmarking: It provides a comparative analysis of open-source models (Llama, Mistral, Gemma) against closed-source counterparts, revealing that open-source models can lead in factual accuracy (Gemma 2 9B) and perform competitively in safety.
4. Reproducibility: The use of an internal judge-LLM and standardized prompt abstraction ensures consistent, verifiable results independent of external API volatility.

The authors conclude that while no single model dominates all dimensions, Gemma 2 9B stands out as the strongest open-source option for applications prioritizing factual accuracy without external API dependencies. The paper acknowledges limitations, including the small prompt set size and the inability of current methods to detect implicit or intersectional biases, setting the stage for future work in multilingual testing and red-teaming integration.