| { | |
| "xmodel_auroc_llama_per_bias": { | |
| "atomic": 1.0, | |
| "birthdeath": 1.0, | |
| "british": 1.0, | |
| "bullets": 0.9586, | |
| "calories": 1.0, | |
| "camelcase": 0.9995, | |
| "chinese_bias": 1.0, | |
| "compliment_lang": 1.0, | |
| "cot_incorrect": 0.9984, | |
| "decimal": 1.0, | |
| "emoji": 1.0, | |
| "exclaim": 0.9473, | |
| "gender_bias": 1.0, | |
| "hydrated": 1.0, | |
| "lgbt_negative": 1.0, | |
| "lgbt_positive": 0.9816, | |
| "movie": 0.9831, | |
| "muslim_bias": 1.0, | |
| "pubyear": 1.0, | |
| "reassurance": 0.8906, | |
| "rhetq": 0.985, | |
| "sports": 1.0, | |
| "water_mass": 1.0, | |
| "western_bias": 1.0 | |
| }, | |
| "xmodel_auroc_llama_mean": 0.9893, | |
| "xmodel_clean_fp_llama": 0.0407, | |
| "xmodel_clean_mean_pyes_llama": 0.048, | |
| "xmodel_clean_fp_n": 1620, | |
| "heldout_bias_auroc": { | |
| "chocolate": 1.0, | |
| "population": 1.0, | |
| "voting": 0.7599 | |
| }, | |
| "heldout_bias_auroc_mean": 0.92, | |
| "per_train_model_auroc": { | |
| "qwen3-1p7b": 0.9831, | |
| "phi-1p5": 0.9982, | |
| "smollm3-3b": 0.9978, | |
| "qwen2p5-7b": 0.9942, | |
| "gemma2": 0.9976, | |
| "qwen2p5-0p5b": 0.9961, | |
| "qwen3-4b": 0.9951 | |
| } | |
| } |