File size: 6,678 Bytes
253d988
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
"""
Random-feature ablation control for the v9c SAE induction-feature ablation result.

The headline finding is: zeroing the top-50 induction-scoring features drops top-1
ICL accuracy from 57.75% -> 47.65% (-10.1pp). Reviewers will fairly ask: what if
50 *random* features did the same thing?

This script:
1. Loads v9c SAE + Gemma-2-2B.
2. Samples 50 *non-induction-cluster* features uniformly at random, for each of
   5 random seeds.
3. Runs the same feature ablation as ablations.py on each random sample.
4. Reports mean accuracy drop across the 5 random samples vs the 10.1pp top-50
   induction drop.

Output: results/random_feature_ablation.json.
"""
import argparse
import json
import random
import time
from functools import partial
from pathlib import Path

import numpy as np
import pandas as pd
import torch

from sae_gemma.model_utils import load_model
from sae_gemma.paths import (
    HOOK_NAME,
    INDUCTION_PROBES_PATH,
    RESULTS_DIR,
    SAE_MAIN_DIR,
    REPO_ROOT,
)

CANDIDATE_IDS_PATH = RESULTS_DIR / "induction_candidate_ids.json"
ABLATION_RESULTS_PATH = RESULTS_DIR / "ablation_results.json"


def load_sae_local(sae_path: Path, device: str):
    from sae_lens.saes.sae import SAE
    sae = SAE.load_from_disk(str(sae_path), device=device)
    sae.eval()
    return sae


@torch.no_grad()
def measure_icl_accuracy(
    model,
    sae,
    probes_df: pd.DataFrame,
    feature_ids_to_ablate: list[int],
    device: str,
    batch_size: int = 16,
) -> float:
    """Run model with SAE-patch + ablation; return top-1 accuracy on induction probes."""

    feature_mask = torch.zeros(sae.cfg.d_sae, dtype=torch.bool, device=device)
    if feature_ids_to_ablate:
        feature_mask[feature_ids_to_ablate] = True

    correct = 0
    total = 0
    probe_tokens = [torch.tensor(t, dtype=torch.long, device=device) for t in probes_df["tokens"]]
    answer_tokens = probes_df["B"].tolist()

    for i in range(0, len(probe_tokens), batch_size):
        batch = probe_tokens[i: i + batch_size]
        max_len = max(len(seq) for seq in batch)
        padded = torch.zeros(len(batch), max_len, dtype=torch.long, device=device)
        seq_lens = []
        for j, seq in enumerate(batch):
            padded[j, : len(seq)] = seq
            seq_lens.append(len(seq))

        def patch_hook(value, hook):
            # value: [batch, seq, d_model]
            b, s, d = value.shape
            flat = value.reshape(b * s, d).float()
            z = sae.encode(flat)                          # [b*s, d_sae]
            z_ablated = z * (~feature_mask).float()       # zero target features
            recon = sae.decode(z_ablated)                 # [b*s, d_model]
            recon_orig = sae.decode(z)
            delta = (recon - recon_orig).reshape(b, s, d).to(value.dtype)
            return value + delta

        logits = model.run_with_hooks(padded, fwd_hooks=[(HOOK_NAME, patch_hook)])

        for j in range(len(batch)):
            final_pos = seq_lens[j] - 1
            pred = logits[j, final_pos].argmax().item()
            if pred == answer_tokens[i + j]:
                correct += 1
            total += 1

    return correct / total


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--n-ablate", type=int, default=50)
    parser.add_argument("--n-seeds", type=int, default=5)
    parser.add_argument("--seed-base", type=int, default=10000)
    parser.add_argument("--batch-size", type=int, default=16)
    parser.add_argument("--device", default="cuda")
    args = parser.parse_args()

    try:
        from dotenv import load_dotenv
        load_dotenv(REPO_ROOT / ".env")
    except ImportError:
        pass

    print("[random_ab] Loading model + SAE ...", flush=True)
    model = load_model(device=args.device)
    sae = load_sae_local(SAE_MAIN_DIR, args.device)
    d_sae = sae.cfg.d_sae

    print(f"[random_ab] Loading induction probes + candidate IDs ...", flush=True)
    probes_df = pd.read_parquet(INDUCTION_PROBES_PATH)
    candidate_ids = set(json.loads(CANDIDATE_IDS_PATH.read_text(encoding="utf-8")))
    non_induction = [i for i in range(d_sae) if i not in candidate_ids]
    print(f"[random_ab] {len(non_induction)} non-induction features available", flush=True)

    # Reference: baseline accuracy + top-50 induction ablation from ablation_results.json
    ablation_results = json.loads(ABLATION_RESULTS_PATH.read_text(encoding="utf-8"))
    baseline_acc = ablation_results["baseline_accuracy"]
    induction_ns = ablation_results["feature_ablation"]["N"]
    induction_accs = ablation_results["feature_ablation"]["accuracy"]
    print(f"[random_ab] Reference baseline: {baseline_acc:.4f}", flush=True)
    print(f"[random_ab] Reference induction ablation: N={induction_ns}  acc={induction_accs}", flush=True)

    # Run random ablation
    results = []
    t0 = time.monotonic()
    for s in range(args.n_seeds):
        seed = args.seed_base + s
        rng = random.Random(seed)
        sample = rng.sample(non_induction, args.n_ablate)
        print(f"[random_ab] Seed {seed}: ablating {args.n_ablate} random non-induction features ...", flush=True)
        acc = measure_icl_accuracy(model, sae, probes_df, sample, args.device, args.batch_size)
        drop = baseline_acc - acc
        elapsed = time.monotonic() - t0
        print(f"[random_ab]   acc={acc:.4f}  drop={drop:+.4f} ({drop*100:+.2f}pp)  elapsed={elapsed/60:.1f}m", flush=True)
        results.append({"seed": seed, "n_ablated": args.n_ablate, "feature_ids": sample[:10], "accuracy": acc, "drop": drop})

    out = {
        "baseline_accuracy": baseline_acc,
        "induction_top50_accuracy": induction_accs[-1] if induction_accs else None,
        "induction_top50_drop": baseline_acc - (induction_accs[-1] if induction_accs else 0.0),
        "random_n_ablated": args.n_ablate,
        "random_n_seeds": args.n_seeds,
        "random_runs": results,
        "random_mean_acc": float(np.mean([r["accuracy"] for r in results])),
        "random_mean_drop": float(np.mean([r["drop"] for r in results])),
        "random_std_drop": float(np.std([r["drop"] for r in results])),
    }
    out_path = RESULTS_DIR / "random_feature_ablation.json"
    out_path.write_text(json.dumps(out, indent=2), encoding="utf-8")
    print(f"\n[random_ab] === RESULTS ===", flush=True)
    print(f"  Top-50 induction features ablated: drop = {out['induction_top50_drop']*100:.2f}pp", flush=True)
    print(f"  50 RANDOM features (mean of {args.n_seeds} seeds): drop = {out['random_mean_drop']*100:.2f}pp ± {out['random_std_drop']*100:.2f}pp", flush=True)
    print(f"  Saved to {out_path}", flush=True)


if __name__ == "__main__":
    main()