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

wisent/comparison/main.py

@@ -0,0 +1,444 @@
+"""
+Comparison of steering methods: Ours vs SAE-based.
+
+This script:
+1. Creates steering vectors using train split of pooled data
+2. Runs base evaluation on test split (no overlap)
+3. Runs steered evaluation on same test split
+"""
+
+from __future__ import annotations
+
+import argparse
+import gc
+import json
+from pathlib import Path
+
+import torch
+from lm_eval import evaluator
+from lm_eval.models.hf_steered import SteeredModel
+
+from wisent.core.models.wisent_model import WisentModel
+from wisent.comparison import ours
+from wisent.comparison import sae
+from wisent.comparison import fgaa
+from wisent.comparison.utils import (
+    load_steering_vector,
+    apply_steering_to_model,
+    remove_steering,
+    convert_to_lm_eval_format,
+    create_test_only_task,
+    extract_accuracy,
+    run_lm_eval_evaluation,
+    run_ll_evaluation,
+)
+
+# Map method names to modules
+METHOD_MODULES = {
+    "caa": ours,
+    "sae": sae,
+    "fgaa": fgaa,
+}
+
+
+def run_single_task(
+    model_name: str,
+    task: str,
+    methods: list[str] = None,
+    num_pairs: int = 50,
+    steering_scales: list[float] = None,
+    device: str = "cuda:0",
+    batch_size: int | str = 1,
+    max_batch_size: int = 8,
+    eval_limit: int | None = None,
+    vectors_dir: Path = None,
+    train_ratio: float = 0.8,
+    caa_layers: str = "12",
+    sae_layers: str = "12",
+    extraction_strategies: list[str] = None,
+    bos_features_source: str = "detected",
+) -> list[dict]:
+    """
+    Run comparison for a single task with multiple methods, scales, and extraction strategies.
+
+    Returns list of result dicts, one per method/scale/strategy combination.
+    """
+    if methods is None:
+        methods = ["caa"]
+    if steering_scales is None:
+        steering_scales = [1.0]
+    if extraction_strategies is None:
+        extraction_strategies = ["mc_balanced"]
+
+    results_list = []
+
+    # Step 1: Create test task
+    test_pct = round((1 - train_ratio) * 100)
+    print(f"\n{'='*60}")
+    print(f"Creating test task for: {task}")
+    print(f"(using {test_pct}% of pooled data)")
+    print(f"{'='*60}")
+
+    task_dict = create_test_only_task(task, train_ratio=train_ratio)
+
+    # Step 2: Generate ALL steering vectors FIRST for ALL strategies (subprocess frees GPU memory after each)
+    # Structure: steering_vectors_data[strategy][method] = steering_data
+    steering_vectors_data = {}
+    train_pct = round(train_ratio * 100)
+
+    for method in methods:
+        if method not in METHOD_MODULES:
+            print(f"WARNING: Method '{method}' not implemented, skipping")
+            continue
+
+        method_module = METHOD_MODULES[method]
+
+        # CAA uses extraction strategy, FGAA/SAE don't
+        for extraction_strategy in (extraction_strategies if method == "caa" else [None]):
+            print(f"\n{'@'*60}")
+            print(f"@ METHOD: {method}, EXTRACTION STRATEGY: {extraction_strategy or 'N/A'}")
+            print(f"{'@'*60}")
+
+            # Select layers based on method: CAA uses caa_layers (default=middle), SAE/FGAA use sae_layers (default=12)
+            method_layers = caa_layers if method == "caa" else sae_layers
+
+            print(f"\n{'='*60}")
+            print(f"Generating steering vector for: {task} (method={method})")
+            print(f"(using {train_pct}% of pooled data - no overlap with test)")
+            print(f"Layers: {method_layers}")
+            print(f"{'='*60}")
+
+            suffix = f"_{extraction_strategy}" if extraction_strategy else ""
+            vector_path = vectors_dir / f"{task}_{method}{suffix}_steering_vector.json"
+
+            kwargs = {
+                "task": task,
+                "model_name": model_name,
+                "output_path": vector_path,
+                "num_pairs": num_pairs,
+                "device": device,
+                "layers": method_layers,
+            }
+            if extraction_strategy:
+                kwargs["extraction_strategy"] = extraction_strategy
+            if method == "fgaa":
+                kwargs["bos_features_source"] = bos_features_source
+
+            method_module.generate_steering_vector(**kwargs)
+
+            steering_data = load_steering_vector(vector_path, default_method=method)
+            if extraction_strategy not in steering_vectors_data:
+                steering_vectors_data[extraction_strategy] = {}
+            steering_vectors_data[extraction_strategy][method] = steering_data
+            print(f"Loaded steering vector with layers: {steering_data['layers']}")
+
+    # Step 3: Load model once for ALL evaluations
+    print(f"\n{'='*60}")
+    print(f"Loading model: {model_name}")
+    print(f"{'='*60}")
+    wisent_model = WisentModel(model_name=model_name, device=device)
+
+    # Step 4: Run base evaluation (no steering applied)
+    print(f"\n{'='*60}")
+    print(f"Running BASE evaluation for: {task}")
+    print(f"{'='*60}")
+
+    base_results = run_lm_eval_evaluation(
+        wisent_model=wisent_model,
+        task_dict=task_dict,
+        task_name=task,
+        batch_size=batch_size,
+        max_batch_size=max_batch_size,
+        limit=eval_limit,
+    )
+    base_acc = extract_accuracy(base_results, task)
+    print(f"Base accuracy (lm-eval): {base_acc:.4f}")
+
+    # Step 4b: Run base LL evaluation (no steering)
+    print(f"\n{'='*60}")
+    print(f"Running BASE LL evaluation for: {task}")
+    print(f"{'='*60}")
+
+    base_ll_acc = run_ll_evaluation(
+        wisent_model=wisent_model,
+        task_dict=task_dict,
+        task_name=task,
+        limit=eval_limit,
+    )
+    print(f"Base accuracy (LL): {base_ll_acc:.4f}")
+
+    # Step 5: Run ALL wisent steered evaluations first (model stays loaded)
+    # Structure: wisent_results[(strategy, method, scale)] = steered_acc
+    wisent_results = {}
+    for method in methods:
+        # CAA uses extraction strategy, FGAA/SAE don't
+        for extraction_strategy in (extraction_strategies if method == "caa" else [None]):
+            if extraction_strategy not in steering_vectors_data:
+                continue
+            if method not in steering_vectors_data[extraction_strategy]:
+                continue
+
+            steering_data = steering_vectors_data[extraction_strategy][method]
+
+            for scale in steering_scales:
+                print(f"\n{'='*60}")
+                print(f"Running STEERED evaluation for: {task} (strategy={extraction_strategy}, method={method}, scale={scale})")
+                print(f"{'='*60}")
+
+                # Apply steering to existing model
+                apply_steering_to_model(wisent_model, steering_data, scale=scale)
+
+                steered_results = run_lm_eval_evaluation(
+                    wisent_model=wisent_model,
+                    task_dict=task_dict,
+                    task_name=task,
+                    batch_size=batch_size,
+                    max_batch_size=max_batch_size,
+                    limit=eval_limit,
+                )
+                steered_acc = extract_accuracy(steered_results, task)
+                print(f"Steered accuracy (lm-eval): {steered_acc:.4f}")
+
+                # Run steered LL evaluation
+                steered_ll_acc = run_ll_evaluation(
+                    wisent_model=wisent_model,
+                    task_dict=task_dict,
+                    task_name=task,
+                    limit=eval_limit,
+                )
+                print(f"Steered accuracy (LL): {steered_ll_acc:.4f}")
+
+                # Remove steering for next iteration
+                remove_steering(wisent_model)
+
+                # Store wisent results
+                wisent_results[(extraction_strategy, method, scale)] = {
+                    "lm_eval": steered_acc,
+                    "ll": steered_ll_acc,
+                }
+
+    # Step 6: Free wisent_model to make room for SteeredModel
+    del wisent_model
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+        torch.cuda.synchronize()
+
+    # Step 7: Run ALL lm-eval native steered evaluations (one at a time)
+    for method in methods:
+        # CAA uses extraction strategy, FGAA/SAE don't
+        for extraction_strategy in (extraction_strategies if method == "caa" else [None]):
+            if extraction_strategy not in steering_vectors_data:
+                continue
+            if method not in steering_vectors_data[extraction_strategy]:
+                continue
+
+            steering_data = steering_vectors_data[extraction_strategy][method]
+
+            for scale in steering_scales:
+                print(f"\n{'='*60}")
+                print(f"Running lm-eval NATIVE steered for: {task} (strategy={extraction_strategy}, method={method}, scale={scale})")
+                print(f"{'='*60}")
+
+                # Convert steering vector to lm-eval format
+                suffix = f"_{extraction_strategy}" if extraction_strategy else ""
+                lm_eval_steer_path = vectors_dir / f"{task}_{method}{suffix}_lm_eval_steer_scale{scale}.pt"
+                convert_to_lm_eval_format(steering_data, lm_eval_steer_path, scale=scale)
+
+                lm_steered = SteeredModel(
+                    pretrained=model_name,
+                    steer_path=str(lm_eval_steer_path),
+                    device=device,
+                    batch_size=batch_size,
+                    max_batch_size=max_batch_size,
+                )
+
+                lm_eval_native_results = evaluator.evaluate(
+                    lm=lm_steered,
+                    task_dict=task_dict,
+                    limit=eval_limit,
+                )
+                lm_eval_native_acc = extract_accuracy(lm_eval_native_results, task)
+                print(f"lm-eval native steered accuracy: {lm_eval_native_acc:.4f}")
+
+                # Clean up SteeredModel to free GPU for next iteration
+                del lm_steered
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+                    torch.cuda.synchronize()
+
+                # Store combined results
+                wisent_result = wisent_results[(extraction_strategy, method, scale)]
+                steered_acc_lm_eval = wisent_result["lm_eval"]
+                steered_acc_ll = wisent_result["ll"]
+                results_list.append({
+                    "task": task,
+                    "extraction_strategy": extraction_strategy or "N/A",
+                    "method": method,
+                    "model": model_name,
+                    "layers": steering_data['layers'],
+                    "num_pairs": num_pairs,
+                    "steering_scale": scale,
+                    "base_accuracy_lm_eval": base_acc,
+                    "base_accuracy_ll": base_ll_acc,
+                    "steered_accuracy_lm_eval": steered_acc_lm_eval,
+                    "steered_accuracy_ll": steered_acc_ll,
+                    "steered_accuracy_lm_eval_native": lm_eval_native_acc,
+                    "difference_lm_eval": steered_acc_lm_eval - base_acc,
+                    "difference_ll": steered_acc_ll - base_ll_acc,
+                    "difference_lm_eval_native": lm_eval_native_acc - base_acc,
+                })
+
+    return results_list
+
+
+def run_comparison(
+    model_name: str,
+    tasks: list[str],
+    methods: list[str] = None,
+    num_pairs: int = 50,
+    steering_scales: list[float] = None,
+    device: str = "cuda:0",
+    batch_size: int | str = 1,
+    max_batch_size: int = 8,
+    eval_limit: int | None = None,
+    output_dir: str = "comparison_results",
+    train_ratio: float = 0.8,
+    caa_layers: str = "12",
+    sae_layers: str = "12",
+    extraction_strategies: list[str] = None,
+    bos_features_source: str = "detected",
+) -> list[dict]:
+    """
+    Run full comparison for multiple tasks, methods, scales, and extraction strategies.
+    """
+    if methods is None:
+        methods = ["caa"]
+    if steering_scales is None:
+        steering_scales = [1.0]
+    if extraction_strategies is None:
+        extraction_strategies = ["mc_balanced"]
+
+    output_dir = Path(output_dir)
+    # Add model name to path (sanitize "/" -> "_")
+    model_dir_name = model_name.replace("/", "_")
+    output_dir = output_dir / model_dir_name
+    vectors_dir = output_dir / "steering_vectors"
+    results_dir = output_dir / "results"
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+    vectors_dir.mkdir(parents=True, exist_ok=True)
+    results_dir.mkdir(parents=True, exist_ok=True)
+
+    all_results = []
+
+    for task in tasks:
+        print(f"\n{'#'*60}")
+        print(f"# TASK: {task}")
+        print(f"{'#'*60}")
+
+        task_results = run_single_task(
+            model_name=model_name,
+            task=task,
+            methods=methods,
+            num_pairs=num_pairs,
+            steering_scales=steering_scales,
+            device=device,
+            batch_size=batch_size,
+            max_batch_size=max_batch_size,
+            eval_limit=eval_limit,
+            vectors_dir=vectors_dir,
+            train_ratio=train_ratio,
+            caa_layers=caa_layers,
+            sae_layers=sae_layers,
+            extraction_strategies=extraction_strategies,
+            bos_features_source=bos_features_source,
+        )
+        all_results.extend(task_results)
+
+        # Save results for this task (includes all strategies)
+        task_results_file = results_dir / f"{task}_results.json"
+        with open(task_results_file, "w") as f:
+            json.dump(task_results, f, indent=2)
+        print(f"Results for {task} saved to: {task_results_file}")
+
+    # Print final summary table
+    print(f"\n{'='*150}")
+    print(f"FINAL COMPARISON RESULTS")
+    print(f"{'='*150}")
+    print(f"Model: {model_name}")
+    print(f"Num pairs: {num_pairs}")
+    print(f"CAA Layers: {caa_layers}")
+    print(f"SAE/FGAA Layers: {sae_layers}")
+    print(f"Strategies: {', '.join(extraction_strategies)}")
+    print(f"{'='*150}")
+    print(f"{'Strategy':<16} {'Task':<10} {'Method':<8} {'Scale':<6} {'Base(E)':<8} {'Base(L)':<8} {'Steer(E)':<9} {'Steer(L)':<9} {'Native':<8} {'Diff(E)':<8} {'Diff(L)':<8} {'Diff(N)':<8}")
+    print(f"{'-'*150}")
+
+    for r in all_results:
+        print(f"{r.get('extraction_strategy', 'N/A'):<16} {r['task']:<10} {r['method']:<8} {r['steering_scale']:<6.1f} "
+              f"{r['base_accuracy_lm_eval']:<8.4f} {r['base_accuracy_ll']:<8.4f} "
+              f"{r['steered_accuracy_lm_eval']:<9.4f} {r['steered_accuracy_ll']:<9.4f} {r['steered_accuracy_lm_eval_native']:<8.4f} "
+              f"{r['difference_lm_eval']:+<8.4f} {r['difference_ll']:+<8.4f} {r['difference_lm_eval_native']:+<8.4f}")
+
+    print(f"{'='*150}")
+
+    print(f"\nSteering vectors saved to: {vectors_dir}")
+    print(f"Results saved to: {results_dir}")
+
+    return all_results
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Compare steering methods")
+    parser.add_argument("--model", default="EleutherAI/gpt-neo-125M", help="Model name")
+    parser.add_argument("--tasks", default="boolq", help="Comma-separated lm-eval tasks (e.g., boolq,cb,copa)")
+    parser.add_argument("--methods", default="caa", help="Comma-separated methods (e.g., caa,sae,fgaa)")
+    parser.add_argument("--num-pairs", type=int, default=50, help="Number of contrastive pairs")
+    parser.add_argument("--scales", default="1.0", help="Comma-separated steering scales (e.g., 0.5,1.0,1.5)")
+    parser.add_argument("--caa-layers", default="12", help="Layer(s) for CAA steering (default: 12)")
+    parser.add_argument("--sae-layers", default="12", help="Layer(s) for SAE/FGAA steering (default: 12)")
+    parser.add_argument("--device", default="cuda:0", help="Device")
+    parser.add_argument("--batch-size", default=1, help="Batch size (int or 'auto')")
+    parser.add_argument("--max-batch-size", type=int, default=8, help="Max batch size for lm-eval internal batching (reduce if OOM)")
+    parser.add_argument("--limit", type=int, default=None, help="Limit eval examples")
+    parser.add_argument("--output-dir", default="wisent/comparison/comparison_results", help="Output directory")
+    parser.add_argument("--train-ratio", type=float, default=0.8, help="Train/test split ratio (default 0.8 = 80%% train, 20%% test)")
+    parser.add_argument("--extraction-strategy", default="mc_balanced",
+                        help="Extraction strategy (comma-separated for multiple). Chat models: chat_mean, chat_first, chat_last, chat_max_norm, chat_weighted, role_play, mc_balanced. Base models: completion_last, completion_mean, mc_completion")
+    parser.add_argument("--bos-features-source", default="detected",
+                        help="BOS features source for FGAA: 'paper' (5 features), 'detected' (12 features), or 'none'")
+
+    args = parser.parse_args()
+
+    # Parse comma-separated values
+    tasks = [t.strip() for t in args.tasks.split(",")]
+    methods = [m.strip() for m in args.methods.split(",")]
+    scales = [float(s.strip()) for s in args.scales.split(",")]
+    extraction_strategies = [s.strip() for s in args.extraction_strategy.split(",")]
+
+    # Parse batch_size (can be int or "auto")
+    batch_size = args.batch_size if args.batch_size == "auto" else int(args.batch_size)
+
+    run_comparison(
+        model_name=args.model,
+        tasks=tasks,
+        methods=methods,
+        num_pairs=args.num_pairs,
+        steering_scales=scales,
+        device=args.device,
+        batch_size=batch_size,
+        max_batch_size=args.max_batch_size,
+        eval_limit=args.limit,
+        output_dir=args.output_dir,
+        train_ratio=args.train_ratio,
+        caa_layers=args.caa_layers,
+        sae_layers=args.sae_layers,
+        extraction_strategies=extraction_strategies,
+        bos_features_source=args.bos_features_source,
+    )
+
+
+if __name__ == "__main__":
+    main()

wisent/comparison/ours.py

@@ -0,0 +1,76 @@
+"""
+Our steering method wrapper for comparison experiments.
+
+Uses the existing wisent infrastructure to create steering vectors.
+Runs steering vector generation in subprocess to guarantee memory cleanup.
+"""
+
+from __future__ import annotations
+
+import json
+import subprocess
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+import torch
+
+from wisent.comparison.utils import apply_steering_to_model, remove_steering, convert_to_lm_eval_format
+
+if TYPE_CHECKING:
+    from wisent.core.models.wisent_model import WisentModel
+
+__all__ = ["generate_steering_vector", "apply_steering_to_model", "remove_steering", "convert_to_lm_eval_format"]
+
+
+def generate_steering_vector(
+    task: str,
+    model_name: str,
+    output_path: str | Path,
+    trait_label: str = "correctness",
+    num_pairs: int = 50,
+    method: str = "caa",
+    layers: str | None = None,
+    normalize: bool = True,
+    device: str = "cuda:0",
+    keep_intermediate: bool = False,
+    extraction_strategy: str = "mc_balanced",
+) -> Path:
+    """
+    Generate a steering vector using wisent CLI in subprocess.
+
+    Runs in subprocess to guarantee GPU memory is freed when done.
+    """
+    output_path = Path(output_path)
+
+    cmd = [
+        "wisent", "generate-vector-from-task",
+        "--task", task,
+        "--trait-label", trait_label,
+        "--model", model_name,
+        "--num-pairs", str(num_pairs),
+        "--method", method,
+        "--output", str(output_path),
+        "--device", device,
+        "--extraction-strategy", extraction_strategy,
+        "--accept-low-quality-vector",
+    ]
+
+    if layers:
+        cmd.extend(["--layers", layers])
+
+    if normalize:
+        cmd.append("--normalize")
+
+    if keep_intermediate:
+        cmd.append("--keep-intermediate")
+
+    result = subprocess.run(cmd)
+
+    if result.returncode != 0:
+        raise RuntimeError(f"Failed to generate steering vector (exit code {result.returncode})")
+
+    return output_path
+
+
+
+