Spaces:
Sleeping
Sleeping
File size: 5,508 Bytes
2aed081 | 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 | import torch
import numpy as np
from sklearn.metrics import f1_score, recall_score, precision_score
# ─────────────────────────────────────────────────────────────────────────────
# Official metric definitions — match AgentHallu paper (Equations 6-11)
# ─────────────────────────────────────────────────────────────────────────────
def step_localization_accuracy(samples, step_preds):
"""
AgentHallu Eq. 11:
Acc_step = |{i ∈ Hhal : pred_step(i) == true_step(i)}| / |Hhal|
Denominator is ALL hallucinated trajectories (regardless of whether
the model's judgment prediction was correct). Do NOT restrict to
TP-only — that would be a lenient variant not comparable to the paper.
"""
correct = total = 0
for s, pred_step in zip(samples, step_preds):
is_hal = s.get("is_hallucination", False)
if isinstance(is_hal, str):
is_hal = is_hal.lower() == "true"
if not is_hal:
continue
total += 1
true_step = s.get("hallucination_step")
if true_step is not None:
true_step = int(true_step)
if pred_step == true_step:
correct += 1
return correct / total if total else 0.0
def tune_threshold(model, val_samples, preprocessor, thresholds=None):
"""
Find the decision threshold on the validation set that maximises
macro-F1. Returns the best threshold and the corresponding metrics.
"""
if thresholds is None:
thresholds = np.arange(0.2, 0.8, 0.02)
model.eval()
device = next(model.parameters()).device
# Collect max-prob per trajectory
max_probs = []
hal_true = []
with torch.no_grad():
for sample in val_samples:
is_hal = sample.get("is_hallucination", False)
if isinstance(is_hal, str):
is_hal = is_hal.lower() == "true"
hal_true.append(int(is_hal))
try:
steps = preprocessor.encode_trajectory(sample)
except Exception:
steps = []
if not steps:
max_probs.append(0.0)
continue
ids = torch.stack([s["encoding"]["input_ids"].squeeze(0) for s in steps]).to(device)
mask = torch.stack([s["encoding"]["attention_mask"].squeeze(0) for s in steps]).to(device)
vocab_size = model.encoder.config.vocab_size
ids = torch.clamp(ids, 0, vocab_size - 1)
logits = model(ids, mask)
probs = torch.sigmoid(logits).cpu().tolist()
if isinstance(probs, float):
probs = [probs]
max_probs.append(max(probs))
best_f1 = -1.0
best_thr = 0.5
for thr in thresholds:
preds = [1 if p > thr else 0 for p in max_probs]
f1 = f1_score(hal_true, preds, average="macro", zero_division=0)
if f1 > best_f1:
best_f1 = f1
best_thr = float(thr)
return best_thr, best_f1
def evaluate(model, test_samples, preprocessor, threshold=0.5):
"""
Full evaluation: judgment (binary) + step localisation.
Uses the supplied threshold (tune it on val, never on test).
"""
model.eval()
device = next(model.parameters()).device
hal_preds = []
hal_true = []
step_preds = []
with torch.no_grad():
for sample in test_samples:
is_hal = sample.get("is_hallucination", False)
if isinstance(is_hal, str):
is_hal = is_hal.lower() == "true"
hal_true.append(int(is_hal))
try:
steps = preprocessor.encode_trajectory(sample)
except Exception:
steps = []
if not steps:
hal_preds.append(0)
step_preds.append(None)
continue
ids = torch.stack([s["encoding"]["input_ids"].squeeze(0) for s in steps]).to(device)
mask = torch.stack([s["encoding"]["attention_mask"].squeeze(0) for s in steps]).to(device)
vocab_size = model.encoder.config.vocab_size
ids = torch.clamp(ids, 0, vocab_size - 1)
logits = model(ids, mask)
probs = torch.sigmoid(logits).cpu().tolist()
if isinstance(probs, float):
probs = [probs]
max_prob = max(probs)
max_prob_idx = probs.index(max_prob)
pred_is_hal = max_prob > threshold
hal_preds.append(int(pred_is_hal))
pred_step = steps[max_prob_idx]["step_idx"] if pred_is_hal else None
step_preds.append(pred_step)
step_acc = step_localization_accuracy(test_samples, step_preds)
macro_f1 = f1_score(hal_true, hal_preds, average="macro", zero_division=0)
macro_rec = recall_score(hal_true, hal_preds, average="macro", zero_division=0)
macro_pre = precision_score(hal_true, hal_preds, average="macro", zero_division=0)
return {
"step_acc": step_acc,
"judgment_f1": macro_f1,
"judgment_recall": macro_rec,
"judgment_precision": macro_pre,
"threshold": threshold,
"n_samples": len(test_samples),
}
|