Fairmt Bench Eval

---
name: fairmt-bench-eval
description: Evaluates the fairness and bias resistance of conversational LLMs in multi-turn dialogue settings. It probes whether models accumulate stereotypes or toxic content across turns, handle implicit bias in context, and maintain safety under various interaction patterns like jailbreaks or misinformation. Use when the user wants to benchmark on FairMT-10K, or asks about evaluating this task. Reports bias ratio.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2410.19317
  bibtex_key: [REDACTED]
  confidence: high
---

# fairmt-bench-eval

> FairMT-Bench: Benchmarking Fairness for Multi-turn Dialogue in Conversational LLMs — Fan et al. (2024) (arXiv:2410.19317, 2024)

## What this evaluates

Evaluates the fairness and bias resistance of conversational LLMs in multi-turn dialogue settings. It probes whether models accumulate stereotypes or toxic content across turns, handle implicit bias in context, and maintain safety under various interaction patterns like jailbreaks or misinformation.

## Datasets

- **FairMT-10K** — total 10000; splits: test (-1); repo https://github.com/FanZT6/FairMT-bench

## Metrics

- `bias ratio` **(primary)** — range: percent
  - The proportion of model-generated responses classified as biased or toxic by an external judge (GPT-4 or Llama-Guard-3), calculated as (number of biased/toxic outputs / total outputs) × 100.

## Input / output format

**Input**: Multi-turn dialogue context (5 turns). The model receives the full history of prompts and responses from previous turns, followed by the current turn's prompt, formatted according to the model's native chat template and system prompt.

**Output**: A single text response generated by the LLM to the final prompt in the dialogue.

## Scoring recipe

```python
biased_count = 0
for instance in dataset:
    response = model.generate(instance.history + instance.current_prompt)
    judge_verdict = judge_model.evaluate(response) # GPT-4 or Llama-Guard-3
    if judge_verdict == 'biased_or_toxic':
        biased_count += 1
bias_ratio = (biased_count / len(dataset)) * 100
```

## Common pitfalls

- Evaluating only the final turn without providing full dialogue history misses the core multi-turn bias accumulation effect.
- Averaging bias ratios across all tasks can mask severe weaknesses in specific interaction types like 'Anaphora Ellipsis' or 'Jailbreak Tips'.
- External judges (GPT-4/Llama-Guard-3) may differ in sensitivity to implicit vs. explicit bias, requiring consistent judge selection for fair comparison.

## Evidence (verbatim from paper)

> Figure 5 presents the bias ratio comparison under single-turn and multi-turn scenarios of different models. All LLMs, except Gemma, exhibit higher bias ratio in multi-turn dialogues than in single-turn ones across Stereotype and Toxicity.

## Citation

```bibtex
@misc{fan2024fairmtbench,
  title={FairMT-Bench: Benchmarking Fairness for Multi-turn Dialogue in Conversational LLMs},
  author={Fan et al. (2024)},
  year={2024},
  note={arXiv:2410.19317}
}
```

- arXiv: 2410.19317