experiment_config_multimodel.json

{
  "experiment": "multi-model intersectional bias study",
  "n_observations": 384,
  "methods": [
    "SEAT",
    "CEAT",
    "MLM"
  ],
  "datasets": [
    "CareerFamily",
    "CompetenceWarmth",
    "PleasantUnpleasant"
  ],
  "pairs": [
    "Hindu_Female",
    "Hindu_Male",
    "LowerCaste_Female",
    "LowerCaste_Male",
    "Muslim_Female",
    "Muslim_Male",
    "UpperCaste_Female",
    "UpperCaste_Male"
  ],
  "n_bootstrap_ceat": 3000,
  "ias_threshold": 0.05,
  "within_method_z_normalisation": true,
  "seed": 42,
  "encoder_models": {
    "SEAT": [
      "mpnet-multi",
      "BGE-M3",
      "E5-large",
      "Qwen3-Emb-4B",
      "IndicSBERT",
      "MuRIL"
    ],
    "CEAT": [
      "mBERT",
      "MuRIL",
      "ModernBERT",
      "XLM-R-large",
      "IndicBERT-v2"
    ],
    "MLM": [
      "mBERT",
      "MuRIL",
      "ModernBERT",
      "XLM-R-large",
      "IndicBERT-v2"
    ]
  },
  "ml_models": [
    "Ridge",
    "RandomForest",
    "GradBoosting",
    "ElasticNet",
    "GaussianProcess",
    "MLP"
  ],
  "best_ml_model": "Ridge",
  "best_lodo_r2": 0.9405090356483208,
  "best_lomo_r2": 0.9386120562963216,
  "best_lomoo_r2": 0.9377524162718508,
  "cv_axes": {
    "LODO": "3 groups (datasets)",
    "LOMO": "3 groups (methods)",
    "LOMOO": "10 groups (encoder models)"
  },
  "feature_cols_real": [
    "score_A",
    "score_B",
    "abs_score_A",
    "abs_score_B",
    "max_single",
    "score_A_times_B",
    "score_A_sq",
    "score_B_sq"
  ],
  "feature_cols_dummy": [
    "dummy_gaussian",
    "dummy_uniform",
    "dummy_permuted_target"
  ],
  "hardware": {
    "gpu": "NVIDIA T4 (15 GB VRAM)",
    "ram": "12.7 GB",
    "platform": "Google Colab free tier"
  },
  "key_findings": {
    "linearity": "Ridge (linear) R\u00b2=0.941 beats all non-linear models",
    "career_family_scale_artifact": "CareerFamily LODO R\u00b2 was negative at 72 rows without normalisation; positive (0.72) at 384 rows with z-norm",
    "snr": 53.6,
    "ablation_p": "< 10^-100 for all 6 conditions",
    "muslim_male_lir_only": {
      "IAS": -0.0470513346414766,
      "LIR": 0.0781596722336495
    },
    "encoder_heterogeneity": {
      "easiest": "mBERT (LOMOO R\u00b2=0.983), XLM-R-large (0.985)",
      "hardest": "Qwen3-Emb-4B (0.429), mpnet-multi (0.539)",
      "interpretation": "SEAT-only models encode bias differently from CEAT/MLM models"
    }
  }
}