wisent 0.7.901__py3-none-any.whl → 0.7.1045__py3-none-any.whl

wisent/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.7.901"
+__version__ = "0.7.1045"
 
 from wisent.core.diversity_processors import (
     OpenerPenaltyProcessor,

wisent/comparison/__init__.py

@@ -0,0 +1 @@
+# Comparison methods for evaluating steering techniques

wisent/comparison/detect_bos_features.py

@@ -0,0 +1,275 @@
+"""
+Detect BOS (Beginning of Sequence) features in Gemma Scope SAEs.
+
+BOS features are SAE features that activate most strongly at the BOS token position.
+These should be filtered out when computing steering vectors as they introduce
+artifacts without contributing to steering.
+
+Reference: "Interpretable Steering of Large Language Models with Feature Guided
+Activation Additions" (arXiv:2501.09929), Appendix G.
+
+Known BOS features from paper (Gemma-2-2B, layer 12, 16k SAE):
+- 11087, 3220, 11752, 12160, 11498
+
+Usage:
+    python -m wisent.comparison.detect_bos_features --model google/gemma-2-2b
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+from datasets import load_dataset
+from tqdm import tqdm
+
+
+# Known BOS feature indices from paper (Appendix G)
+KNOWN_BOS_FEATURES = {
+    "google/gemma-2-2b": [11087, 3220, 11752, 12160, 11498],
+    "google/gemma-2-9b": [],  # Not listed in paper
+}
+
+
+def load_sample_texts(num_samples: int = 2000, min_length: int = 50) -> list[str]:
+    """Load sample texts from WikiText dataset."""
+    print(f"Loading up to {num_samples} sample texts from WikiText...")
+
+    dataset = load_dataset("wikitext", "wikitext-103-v1", split="train")
+
+    texts = []
+    for item in dataset:
+        if len(texts) >= num_samples:
+            break
+        text = item["text"].strip()
+        if len(text) >= min_length:
+            texts.append(text)
+
+    print(f"   Loaded {len(texts)} texts")
+    return texts
+
+
+def detect_bos_features(
+    model,
+    tokenizer,
+    sae,
+    layer_idx: int,
+    device: str,
+    texts: list[str],
+    top_k: int = 10,
+    batch_size: int = 8,
+) -> tuple[list[int], dict[str, torch.Tensor]]:
+    """
+    Detect BOS features by finding features that activate most strongly at position 0.
+
+    Computes statistics (mean, variance, median) of activation at BOS position for each
+    SAE feature across all samples, then returns the top-k features with highest mean
+    BOS activation.
+
+    Reference: FGAA paper (arXiv:2501.09929) identifies BOS features as those that
+    "exclusively had the strongest activation on the BOS token".
+
+    Args:
+        model: HuggingFace model
+        tokenizer: Tokenizer
+        sae: SAE object from sae_lens
+        layer_idx: Layer index (0-indexed)
+        device: Device
+        texts: List of sample texts to analyze
+        top_k: Number of top BOS features to return (default 10)
+        batch_size: Batch size for processing
+
+    Returns:
+        Tuple of (top_bos_feature_indices, stats_dict) where stats_dict contains
+        'mean', 'variance', and 'median' tensors of shape [d_sae].
+    """
+    d_sae = sae.cfg.d_sae
+
+    # Collect all BOS activations on CPU for stable statistics computation
+    all_bos_activations = []
+
+    print(f"Detecting BOS features from {len(texts)} samples...")
+    print(f"   Layer: {layer_idx}, d_sae: {d_sae}")
+
+    # Use hook to capture only the layer we need (not all 26 layers)
+    captured_acts = {}
+    def capture_hook(module, input, output):
+        captured_acts["hidden"] = output[0].detach()
+
+    # Register hook on the specific layer
+    target_layer = model.model.layers[layer_idx]
+    hook_handle = target_layer.register_forward_hook(capture_hook)
+
+    try:
+        for i in tqdm(range(0, len(texts), batch_size), desc="Processing"):
+            batch_texts = texts[i:i + batch_size]
+
+            inputs = tokenizer(
+                batch_texts,
+                return_tensors="pt",
+                truncation=True,
+                max_length=128,
+                padding=True,
+            )
+            inputs = {k: v.to(device) for k, v in inputs.items()}
+
+            with torch.no_grad():
+                # Use model.model to skip lm_head (saves ~66MB per forward)
+                model.model(**inputs, use_cache=False)
+
+            # Get captured hidden states (only this layer, not all 26)
+            acts = captured_acts["hidden"].to(sae.W_enc.dtype)
+            latents = sae.encode(acts)
+
+            attention_mask = inputs.get("attention_mask")
+            for j in range(latents.shape[0]):
+                seq_len = int(attention_mask[j].sum().item()) if attention_mask is not None else latents.shape[1]
+                sample_latents = latents[j, :seq_len, :]  # [seq_len, d_sae]
+
+                # Collect BOS activation (position 0) - move to CPU immediately
+                bos_act = sample_latents[0].float().cpu()
+                all_bos_activations.append(bos_act)
+    finally:
+        hook_handle.remove()
+
+    # Compute statistics (all on CPU)
+    # Stack all activations for stable computation
+    all_bos_tensor = torch.stack(all_bos_activations, dim=0)  # [num_samples, d_sae]
+    mean_bos = all_bos_tensor.mean(dim=0)
+    variance_bos = all_bos_tensor.var(dim=0)
+    median_bos = all_bos_tensor.median(dim=0).values
+
+    stats = {
+        "mean": mean_bos,
+        "variance": variance_bos,
+        "median": median_bos,
+    }
+
+    # Select top features by BOS activation
+    top_indices = mean_bos.topk(top_k).indices.tolist()
+
+    print(f"\nDetected top {top_k} BOS features by mean activation")
+    return top_indices, stats
+
+
+def compare_with_known(model_name: str, detected: list[int], stats: dict[str, torch.Tensor]) -> None:
+    """Compare detected BOS features with known list from paper."""
+    mean_bos = stats["mean"]
+    variance_bos = stats["variance"]
+    median_bos = stats["median"]
+
+    known = KNOWN_BOS_FEATURES.get(model_name, [])
+    detected_set = set(detected)
+    known_set = set(known)
+
+    print(f"\n{'='*60}")
+    print(f"BOS Feature Comparison for {model_name}")
+    print(f"{'='*60}")
+    print(f"Known (paper): {sorted(known_set)}")
+    print(f"Detected:      {sorted(detected_set)}")
+    print(f"{'='*60}")
+    print(f"Common:        {sorted(detected_set & known_set)}")
+    print(f"Only in paper: {sorted(known_set - detected_set)}")
+    print(f"Only detected: {sorted(detected_set - known_set)}")
+
+    if known:
+        print(f"\nKnown features - BOS activation stats:")
+        for idx in sorted(known):
+            mean_val = mean_bos[idx].item()
+            var_val = variance_bos[idx].item()
+            median_val = median_bos[idx].item()
+            status = "detected" if idx in detected_set else "missed"
+            print(f"   Feature {idx}: mean={mean_val:.4f}, var={var_val:.4f}, median={median_val:.4f} ({status})")
+
+    print(f"\nTop 20 features by mean BOS activation:")
+    for rank, idx in enumerate(mean_bos.topk(20).indices.tolist(), 1):
+        mean_val = mean_bos[idx].item()
+        var_val = variance_bos[idx].item()
+        median_val = median_bos[idx].item()
+        marker = " (known)" if idx in known_set else ""
+        print(f"   {rank:2}. Feature {idx}: mean={mean_val:.4f}, var={var_val:.4f}, median={median_val:.4f}{marker}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Detect BOS features in Gemma Scope SAEs")
+    parser.add_argument("--model", default="google/gemma-2-2b", help="Model name")
+    parser.add_argument("--layer", type=int, default=12, help="Layer index")
+    parser.add_argument("--num-samples", type=int, default=1000, help="Number of text samples")
+    parser.add_argument("--top-k", type=int, default=20, help="Number of top BOS features to detect")
+    parser.add_argument("--batch-size", type=int, default=1, help="Batch size")
+    parser.add_argument("--device", default="cuda:0", help="Device")
+    parser.add_argument("--output-dir", default="wisent/comparison/results", help="Output directory")
+    args = parser.parse_args()
+
+    print(f"Model: {args.model}")
+    print(f"Layer: {args.layer}")
+    print(f"Top-k: {args.top_k}")
+
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+
+    print(f"\nLoading model...")
+    model = AutoModelForCausalLM.from_pretrained(
+        args.model,
+        torch_dtype=torch.bfloat16,
+        device_map=args.device,
+        trust_remote_code=True,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+    model.eval()
+
+    print(f"\nLoading SAE...")
+    from sae_lens import SAE
+
+    release = "gemma-scope-2b-pt-res-canonical" if "2b" in args.model.lower() else "gemma-scope-9b-pt-res-canonical"
+    sae_id = f"layer_{args.layer}/width_16k/canonical"
+    sae, _, _ = SAE.from_pretrained(release=release, sae_id=sae_id, device=args.device)
+
+    texts = load_sample_texts(args.num_samples)
+
+    bos_features, stats = detect_bos_features(
+        model=model,
+        tokenizer=tokenizer,
+        sae=sae,
+        layer_idx=args.layer,
+        device=args.device,
+        texts=texts,
+        top_k=args.top_k,
+        batch_size=args.batch_size,
+    )
+
+    compare_with_known(args.model, bos_features, stats)
+
+    # Build detected features with their stats
+    detected_with_stats = [
+        {
+            "feature": idx,
+            "mean_bos_activation": stats["mean"][idx].item(),
+            "variance_bos_activation": stats["variance"][idx].item(),
+            "median_bos_activation": stats["median"][idx].item(),
+        }
+        for idx in bos_features
+    ]
+
+    output = {
+        "model": args.model,
+        "layer": args.layer,
+        "top_k": args.top_k,
+        "num_samples": len(texts),
+        "detected_bos_features": detected_with_stats,
+        "known_bos_features": KNOWN_BOS_FEATURES.get(args.model, []),
+    }
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    output_path = output_dir / f"bos_features_{args.model.replace('/', '_')}_layer{args.layer}.json"
+    with open(output_path, "w") as f:
+        json.dump(output, f, indent=2)
+    print(f"\nSaved to {output_path}")
+
+
+if __name__ == "__main__":
+    main()