models.yaml

# ─────────────────────────────────────────────────────────────────────────────
# Knowledge Drift Detection — Model Registry
# ─────────────────────────────────────────────────────────────────────────────
# All model-specific configuration in one place.
# Adding a new model = adding ~8 lines here. No Python edits needed.

models:
  llama2:
    name: "meta-llama/Llama-2-7b-hf"
    hidden_dim: 4096
    n_layers: 32
    cutoff: "2022-09-01"
    cutoff_months: 30          # months since Jan 2020
    drift_key: "is_drifted_llama2"
    is_instruct: false
    notes: "Base model, no chat template. Use raw completion."

  mistral:
    name: "mistralai/Mistral-7B-Instruct-v0.2"
    hidden_dim: 4096
    n_layers: 32
    cutoff: "2023-12-01"
    cutoff_months: 47
    drift_key: "is_drifted_mistral"
    is_instruct: true
    notes: "Instruct model with [INST] template."

  llama31:
    name: "meta-llama/Llama-3.1-8B-Instruct"
    hidden_dim: 4096
    n_layers: 32
    cutoff: "2023-12-01"
    cutoff_months: 47
    drift_key: "is_drifted_llama31"
    is_instruct: true
    notes: "Instruct model. Requires HF token for gated access."

  qwen25:
    name: "Qwen/Qwen2.5-7B-Instruct"
    hidden_dim: 3584
    n_layers: 28
    cutoff: "2023-12-31"
    cutoff_months: 48
    drift_key: "is_drifted_qwen25"
    is_instruct: true
    notes: "Primary development model. Cache already exists."

  gemma2:
    name: "google/gemma-2-9b-it"
    hidden_dim: 3584
    n_layers: 42
    cutoff: "2024-06-01"
    cutoff_months: 54
    drift_key: "is_drifted_gemma2"
    is_instruct: true
    notes: "Newest cutoff. Fewest drifted samples (1178)."

# ─────────────────────────────────────────────────────────────────────────────
# Experiment defaults
# ─────────────────────────────────────────────────────────────────────────────

defaults:
  dataset: "data/knowledge_drift_unified_tier1.json"
  output_dir: "data/experiments/v4"
  probe_device: "cuda:0"
  max_iter: 2000
  cv_max_iter: 500
  n_permutations: 1000
  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]
  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]
  cv_folds: 3
  batch_size: 1              # for extraction (memory safety)
  max_length: 512            # tokenizer truncation