Raniahossam33/knowledge-drift-experiments

1# ─────────────────────────────────────────────────────────────────────────────2# Knowledge Drift Detection — Model Registry3# ─────────────────────────────────────────────────────────────────────────────4# All model-specific configuration in one place.5# Adding a new model = adding ~8 lines here. No Python edits needed.6 7models:8  llama2:9    name: "meta-llama/Llama-2-7b-hf"10    hidden_dim: 409611    n_layers: 3212    cutoff: "2022-09-01"13    cutoff_months: 30          # months since Jan 202014    drift_key: "is_drifted_llama2"15    is_instruct: false16    notes: "Base model, no chat template. Use raw completion."17 18  mistral:19    name: "mistralai/Mistral-7B-Instruct-v0.2"20    hidden_dim: 409621    n_layers: 3222    cutoff: "2023-12-01"23    cutoff_months: 4724    drift_key: "is_drifted_mistral"25    is_instruct: true26    notes: "Instruct model with [INST] template."27 28  llama31:29    name: "meta-llama/Llama-3.1-8B-Instruct"30    hidden_dim: 409631    n_layers: 3232    cutoff: "2023-12-01"33    cutoff_months: 4734    drift_key: "is_drifted_llama31"35    is_instruct: true36    notes: "Instruct model. Requires HF token for gated access."37 38  qwen25:39    name: "Qwen/Qwen2.5-7B-Instruct"40    hidden_dim: 358441    n_layers: 2842    cutoff: "2023-12-31"43    cutoff_months: 4844    drift_key: "is_drifted_qwen25"45    is_instruct: true46    notes: "Primary development model. Cache already exists."47 48  gemma2:49    name: "google/gemma-2-9b-it"50    hidden_dim: 358451    n_layers: 4252    cutoff: "2024-06-01"53    cutoff_months: 5454    drift_key: "is_drifted_gemma2"55    is_instruct: true56    notes: "Newest cutoff. Fewest drifted samples (1178)."57 58# ─────────────────────────────────────────────────────────────────────────────59# Experiment defaults60# ─────────────────────────────────────────────────────────────────────────────61 62defaults:63  dataset: "data/knowledge_drift_unified_tier1.json"64  output_dir: "data/experiments/v4"65  probe_device: "cuda:0"66  max_iter: 200067  cv_max_iter: 50068  n_permutations: 100069  lambda_grid: [1.0e-5, 5.0e-5, 1.0e-4, 5.0e-4, 1.0e-3, 5.0e-3, 1.0e-2, 5.0e-2]70  sparsity_lambdas: [1.0e-6, 1.0e-5, 5.0e-5, 1.0e-4, 5.0e-4, 1.0e-3, 5.0e-3, 1.0e-2, 5.0e-2, 0.1, 0.2]71  cv_folds: 372  batch_size: 1              # for extraction (memory safety)73  max_length: 512            # tokenizer truncation