Spaces:
Running
Running
File size: 6,762 Bytes
dfbca8a 573a89c a202ba5 573a89c a202ba5 573a89c a202ba5 573a89c a202ba5 573a89c dfbca8a 573a89c dfbca8a 573a89c dfbca8a 573a89c dfbca8a 573a89c dfbca8a 573a89c |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 |
import plotly.graph_objects as go
import streamlit as st
import torch
import torch.nn.functional as F
from datasets import load_dataset
from pydantic import BaseModel
from rich.progress import track
from sklearn.metrics import roc_auc_score, roc_curve
from transformers import AutoModelForSequenceClassification, AutoTokenizer
class DatasetArgs(BaseModel):
dataset_address: str
train_dataset_range: int
test_dataset_range: int
class LlamaGuardFineTuner:
def __init__(self, streamlit_mode: bool = False):
self.streamlit_mode = streamlit_mode
def load_dataset(self, dataset_args: DatasetArgs):
dataset = load_dataset(dataset_args.dataset_address)
self.train_dataset = (
dataset["train"]
if dataset_args.train_dataset_range <= 0
or dataset_args.train_dataset_range > len(dataset["train"])
else dataset["train"].select(range(dataset_args.train_dataset_range))
)
self.test_dataset = (
dataset["test"]
if dataset_args.test_dataset_range <= 0
or dataset_args.test_dataset_range > len(dataset["test"])
else dataset["test"].select(range(dataset_args.test_dataset_range))
)
def load_model(self, model_name: str = "meta-llama/Prompt-Guard-86M"):
self.device = "cuda" if torch.cuda.is_available() else "cpu"
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
self.model = AutoModelForSequenceClassification.from_pretrained(model_name).to(
self.device
)
def show_dataset_sample(self):
if self.streamlit_mode:
st.markdown("### Train Dataset Sample")
st.dataframe(self.train_dataset.to_pandas().head())
st.markdown("### Test Dataset Sample")
st.dataframe(self.test_dataset.to_pandas().head())
def evaluate_batch(
self,
texts,
batch_size: int = 32,
positive_label: int = 2,
temperature: float = 1.0,
truncation: bool = True,
max_length: int = 512,
) -> list[float]:
self.model.eval()
encoded_texts = self.tokenizer(
texts,
padding=True,
truncation=truncation,
max_length=max_length,
return_tensors="pt",
)
dataset = torch.utils.data.TensorDataset(
encoded_texts["input_ids"], encoded_texts["attention_mask"]
)
data_loader = torch.utils.data.DataLoader(dataset, batch_size=batch_size)
scores = []
progress_bar = (
st.progress(0, text="Evaluating") if self.streamlit_mode else None
)
for i, batch in track(
enumerate(data_loader), description="Evaluating", total=len(data_loader)
):
input_ids, attention_mask = [b.to(self.device) for b in batch]
with torch.no_grad():
logits = self.model(
input_ids=input_ids, attention_mask=attention_mask
).logits
scaled_logits = logits / temperature
probabilities = F.softmax(scaled_logits, dim=-1)
positive_class_probabilities = (
probabilities[:, positive_label].cpu().numpy()
)
scores.extend(positive_class_probabilities)
if progress_bar:
progress_percentage = (i + 1) * 100 // len(data_loader)
progress_bar.progress(
progress_percentage,
text=f"Evaluating batch {i + 1}/{len(data_loader)}",
)
return scores
def visualize_roc_curve(self, test_scores: list[float]):
test_labels = [int(elt) for elt in self.test_dataset["label"]]
fpr, tpr, _ = roc_curve(test_labels, test_scores)
roc_auc = roc_auc_score(test_labels, test_scores)
fig = go.Figure()
fig.add_trace(
go.Scatter(
x=fpr,
y=tpr,
mode="lines",
name=f"ROC curve (area = {roc_auc:.3f})",
line=dict(color="darkorange", width=2),
)
)
fig.add_trace(
go.Scatter(
x=[0, 1],
y=[0, 1],
mode="lines",
name="Random Guess",
line=dict(color="navy", width=2, dash="dash"),
)
)
fig.update_layout(
title="Receiver Operating Characteristic",
xaxis_title="False Positive Rate",
yaxis_title="True Positive Rate",
xaxis=dict(range=[0.0, 1.0]),
yaxis=dict(range=[0.0, 1.05]),
legend=dict(x=0.8, y=0.2),
)
if self.streamlit_mode:
st.plotly_chart(fig)
else:
fig.show()
def visualize_score_distribution(self, scores: list[float]):
test_labels = [int(elt) for elt in self.test_dataset["label"]]
positive_scores = [scores[i] for i in range(500) if test_labels[i] == 1]
negative_scores = [scores[i] for i in range(500) if test_labels[i] == 0]
fig = go.Figure()
# Plotting positive scores
fig.add_trace(
go.Histogram(
x=positive_scores,
histnorm="probability density",
name="Positive",
marker_color="darkblue",
opacity=0.75,
)
)
# Plotting negative scores
fig.add_trace(
go.Histogram(
x=negative_scores,
histnorm="probability density",
name="Negative",
marker_color="darkred",
opacity=0.75,
)
)
# Updating layout
fig.update_layout(
title="Score Distribution for Positive and Negative Examples",
xaxis_title="Score",
yaxis_title="Density",
barmode="overlay",
legend_title="Scores",
)
# Display the plot
if self.streamlit_mode:
st.plotly_chart(fig)
else:
fig.show()
def evaluate_model(
self,
batch_size: int = 32,
positive_label: int = 2,
temperature: float = 3.0,
truncation: bool = True,
max_length: int = 512,
):
test_scores = self.evaluate_batch(
self.test_dataset["text"],
batch_size=batch_size,
positive_label=positive_label,
temperature=temperature,
truncation=truncation,
max_length=max_length,
)
self.visualize_roc_curve(test_scores)
self.visualize_score_distribution(test_scores)
return test_scores
|