File size: 2,872 Bytes
829572e 5094855 829572e 5094855 829572e 5094855 685f359 5094855 685f359 5094855 7477097 5094855 7477097 5094855 |
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 |
# metrics.py
from model_loader import classifier_model, metrics_models
import torch
import numpy as np
import time
def softmax(logits):
exp_logits = np.exp(logits - np.max(logits))
return exp_logits / exp_logits.sum()
def compute_reward_scores(original, paraphrase):
"""
Compute reward scores for a paraphrased comment.
Returns a dictionary with empathy, toxicity, bias, hallucination, and reward scores.
"""
try:
start_time = time.time()
print("Starting reward computation...")
# Check if paraphrase is valid
if not isinstance(paraphrase, str) or "Error: Unable to generate paraphrase" in paraphrase:
print(f"Invalid paraphrase: {paraphrase}. Returning default scores.")
return {
"empathy": 0.0,
"toxicity": 1.0,
"bias": 1.0,
"hallucination": 1.0,
"reward": 0.0
}
# Classify the paraphrased comment
print("Starting classification...")
inputs = classifier_model.tokenizer(
paraphrase,
return_tensors="pt",
truncation=True,
padding=True,
max_length=512
).to(classifier_model.device)
with torch.no_grad():
outputs = classifier_model.model(**inputs)
logits = outputs.logits.cpu().numpy()[0]
probs = softmax(logits)
toxicity = probs[1] # Assuming label 1 is toxic
empathy = 1.0 - toxicity # Simplified empathy score
bias = probs[1] # Placeholder for bias
print(f"Classification took {time.time() - start_time:.2f} seconds")
# Compute semantic similarity using Sentence-BERT
print("Computing semantic similarity...")
sentence_bert = metrics_models.sentence_bert
embeddings = sentence_bert.encode([original, paraphrase], convert_to_tensor=True)
similarity = torch.cosine_similarity(embeddings[0], embeddings[1], dim=0).item()
hallucination = 1.0 - similarity # High difference means potential hallucination
print(f"Semantic similarity computed: {similarity}")
# Compute reward score (weighted combination)
reward = 0.4 * empathy - 0.2 * toxicity - 0.2 * bias - 0.2 * hallucination
reward = max(0.0, min(1.0, reward))
print(f"Total processing time: {time.time() - start_time:.2f} seconds")
return {
"empathy": empathy,
"toxicity": toxicity,
"bias": bias,
"hallucination": hallucination,
"reward": reward
}
except Exception as e:
print(f"Error in reward computation: {str(e)}")
return {
"empathy": 0.0,
"toxicity": 1.0,
"bias": 1.0,
"hallucination": 1.0,
"reward": 0.0
} |