Bias Evaluation on Male-biased prompts
0.53Male Bias (Base)Manually curated
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Manually curated2025.12 | 0.53 | 0.49 | 0.31 | 96 | 96 | 96 | 0 | 0 | 0 | |
| LLMBackbone LLM=Llama 3.2 1B2025.12 | 0.53 | 0.59 | 0.42 | 54 | 54 | 50 | 1 | 1 | 2 | |
| LLMBackbone LLM=Qwen3 8B2025.12 | 0.55 | 0.57 | 0.35 | 53 | 53 | 53 | 1 | 1 | 0 | |
| LLM (biased)Backbone LLM=Llama 3.2 1B2025.12 | 0.61 | 0.63 | 0.42 | 55 | 55 | 55 | 10 | 5 | 11 | |
| LLM (biased)Backbone LLM=Qwen3 8B2025.12 | 0.63 | 0.65 | 0.41 | 58 | 58 | 58 | 1 | 2 | 2 | |
| BGPSBackbone LLM=Qwen3 8B, lambda=102025.12 | 0.66 | 0.56 | 0.48 | 53 | 52 | 60 | 4 | 2 | 0 | |
| LLM (biased)Backbone LLM=Mistral 7B 0.22025.12 | 0.67 | 0.66 | 0.43 | 60 | 60 | 60 | 0 | 0 | 0 | |
| BGPSBackbone LLM=Llama 3.2 1B, lambda=102025.12 | 0.67 | 0.62 | 0.52 | 51 | 53 | 47 | 3 | 1 | 9 | |
| LLMBackbone LLM=Mistral 7B 0.22025.12 | 0.69 | 0.59 | 0.44 | 71 | 71 | 50 | 1 | 1 | 1 | |
| BGPSBackbone LLM=Mistral 7B 0.2, lambda=102025.12 | 0.76 | 0.66 | 0.46 | 52 | 51 | 56 | 2 | 0 | 3 | |
| PEZ2025.12 | 0.8 | 0.78 | 0.84 | 1,387 | 2,703 | 1,602 | 94 | 94 | 97 | |
| BGPSBackbone LLM=Qwen3 8B, lambda=1002025.12 | 0.85 | 0.66 | 0.54 | 78 | 56 | 107 | 15 | 8 | 12 | |
| BGPSBackbone LLM=Llama 3.2 1B, lambda=1002025.12 | 0.85 | 0.68 | 0.79 | 70 | 52 | 67 | 39 | 21 | 42 | |
| BGPSBackbone LLM=Mistral 7B 0.2, lambda=1002025.12 | 0.92 | 0.79 | 0.7 | 122 | 90 | 160 | 17 | 22 | 10 |