File size: 3,952 Bytes
60b21d3 | 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 | # SPDX-FileCopyrightText: 2025 Stanford University, ETH Zurich, and the project authors (see CONTRIBUTORS.md)
# SPDX-FileCopyrightText: 2025 This source file is part of the OpenTSLM open-source project.
#
# SPDX-License-Identifier: MIT
import re
import sys
import io
import base64
from typing import Dict, Any
import matplotlib.pyplot as plt
from common_evaluator_plot import CommonEvaluatorPlot
from opentslm.time_series_datasets.TSQADataset import TSQADataset
def evaluate_tsqa(ground_truth: str, prediction: str) -> Dict[str, Any]:
"""
Evaluate TSQA predictions against ground truth.
- TSQA answers are in the format "(a)", "(b)", or "(c)".
- The model might answer with a longer explanation like
"(a) This time series has a constant volatility.", so we compare only
the first 3 characters of the answer (e.g., "(a)").
- If the prediction contains an "Answer:" tag, take everything after it.
"""
# Clean up strings for comparison
gt_clean = ground_truth.lower().strip()
pred_clean = prediction.lower().strip()
# Only compare the first 3 characters of the ground truth and prediction
gt_clean, pred_clean = gt_clean[:3], pred_clean[:3]
# Extract the actual answer from the prediction (everything after "Answer:")
answer_match = re.search(r"answer:\s*(.+)", pred_clean, re.IGNORECASE)
if answer_match:
pred_answer = answer_match.group(1).strip()
else:
# If no "Answer:" found, use the entire prediction
pred_answer = pred_clean
# Calculate accuracy (exact match)
accuracy = int(gt_clean == pred_answer)
return {
"accuracy": accuracy,
}
def generate_time_series_plot(time_series) -> str:
"""
Create a base64 PNG plot from one or more time series.
- TSQA typically has a single 1D time series, but we handle multiple for robustness.
- Accepts either a single 1D list/array or a list of 1D lists/arrays.
"""
if time_series is None:
return None
# Ensure we have a list of series
ts_list = list(time_series)
# If it's a flat list of numbers, wrap it as a single series
if ts_list and all(isinstance(x, (int, float)) for x in ts_list):
ts_list = [ts_list]
num_series = len(ts_list)
fig, axes = plt.subplots(num_series, 1, figsize=(10, 4 * num_series), sharex=True)
if num_series == 1:
axes = [axes]
for i, series in enumerate(ts_list):
axes[i].plot(series, marker="o", linestyle="-", markersize=0)
axes[i].grid(True, alpha=0.3)
axes[i].set_title(f"Time Series {i + 1}")
axes[i].set_ylabel("Value")
axes[-1].set_xlabel("Time")
plt.tight_layout()
img_buffer = io.BytesIO()
plt.savefig(img_buffer, format="png", bbox_inches="tight", dpi=100)
plt.close()
img_buffer.seek(0)
image_data = base64.b64encode(img_buffer.getvalue()).decode("utf-8")
return image_data
def main():
"""Main function to run TSQA evaluation with plotting."""
if len(sys.argv) != 2:
print("Usage: python evaluate_tsqa_plot.py <model_name>")
print("Example: python evaluate_tsqa_plot.py openai-gpt-4o")
sys.exit(1)
model_name = sys.argv[1]
dataset_classes = [TSQADataset]
evaluation_functions = {
"TSQADataset": evaluate_tsqa,
}
evaluator = CommonEvaluatorPlot()
plot_functions = {
"TSQADataset": generate_time_series_plot,
}
results_df = evaluator.evaluate_multiple_models(
model_names=[model_name],
dataset_classes=dataset_classes,
evaluation_functions=evaluation_functions,
plot_functions=plot_functions,
max_samples=None, # Set to None for full evaluation
max_new_tokens=80,
)
print("\n" + "=" * 80)
print("FINAL RESULTS SUMMARY")
print("=" * 80)
print(results_df.to_string(index=False))
return results_df
if __name__ == "__main__":
main()
|