| """ |
| Stage 02 (v8b): Build per-layer direction subspaces (dense, no MoE). |
| |
| For each captured layer: |
| - mean-difference vector (high-reflection minus low-reflection class) |
| - PCA-denoise within the top-N PCs of all activations |
| - orthogonalize against the layer's general mean |
| |
| No model load is needed here — Qwen3-8B is dense, so there is no expert |
| weight mask to compute. We work directly off the stage-01 activations. |
| |
| Resume: skip if DIRECTIONS already exists. |
| """ |
| import argparse, os, sys |
| sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) |
|
|
| import torch |
| from configs import get_config |
| from configs.paths import LOG_DIR, dim_paths, ensure_dirs |
| from src.directions import build_layer_directions |
| from src.utils import setup_logger, write_json |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--dimension", default="monitoring") |
| ap.add_argument("--n-pca-components", type=int, default=100) |
| ap.add_argument("--min-residual-after-general", type=float, default=0.20) |
| ap.add_argument("--disable-pca", action="store_true", |
| help="Ablation: skip PCA denoising on mean-diff.") |
| ap.add_argument("--disable-ortho", action="store_true", |
| help="Ablation: skip orthogonalization vs general mean.") |
| ap.add_argument("--force", action="store_true") |
| args = ap.parse_args() |
|
|
| ensure_dirs(args.dimension) |
| cfg = get_config(args.dimension) |
| p = dim_paths(args.dimension) |
|
|
| log = setup_logger("02_directions", |
| os.path.join(LOG_DIR, f"02_directions_{cfg.NAME}.log")) |
| log.info("=" * 70) |
| log.info(f"Stage 02 [{cfg.NAME}] (v8b dense) pca_n={args.n_pca_components}") |
| log.info("=" * 70) |
|
|
| if os.path.exists(p.DIRECTIONS) and not args.force: |
| try: |
| ex = torch.load(p.DIRECTIONS, map_location="cpu", weights_only=False) |
| log.info(f" [resume] {p.DIRECTIONS} exists " |
| f"({len(ex.get('directions', {}))} layers) — SKIP. " |
| f"Use --force to recompute.") |
| return |
| except Exception as e: |
| log.warning(f" [resume] unreadable ({e}); recomputing") |
|
|
| if not os.path.exists(p.ACTIVATIONS): |
| log.error(f"missing {p.ACTIVATIONS} — run stage 01 first") |
| sys.exit(1) |
|
|
| blob = torch.load(p.ACTIVATIONS, map_location="cpu", weights_only=False) |
| per_layer = blob["per_layer"] |
|
|
| log.info("PCA-denoised mean-diff + ortho-vs-general per layer...") |
| per_layer_data = { |
| L: {"acts": per_layer[L]["acts"], "labels": per_layer[L]["labels"]} |
| for L in per_layer |
| } |
| directions, diagnostics = build_layer_directions( |
| per_layer_data, |
| n_pca_components=args.n_pca_components, |
| min_residual_after_general=args.min_residual_after_general, |
| disable_pca=args.disable_pca, |
| disable_ortho=args.disable_ortho, |
| logger=log, |
| ) |
|
|
| save = { |
| "dimension": cfg.NAME, |
| "n_pca_components": args.n_pca_components, |
| "directions": directions, |
| "diagnostics": diagnostics, |
| "target_layers": cfg.TARGET_LAYERS, |
| } |
| tmp = p.DIRECTIONS + ".tmp" |
| torch.save(save, tmp) |
| os.replace(tmp, p.DIRECTIONS) |
| log.info(f"Saved {p.DIRECTIONS} ({len(directions)} layers kept). Done.") |
| write_json( |
| {"kept": sorted(directions.keys()), "diagnostics": diagnostics}, |
| p.DIRECTIONS.replace(".pt", "_summary.json"), |
| ) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
