import json
import plotly.graph_objects as go
from typing import Dict, List, Tuple
from collections import defaultdict
def create_distribution_plots(metrics_data: Dict[str, List[float]]) -> List[str]:
"""Create distribution histogram plots for each metric."""
figures = []
for metric, scores in metrics_data.items():
fig = go.Figure()
fig.add_trace(go.Histogram(x=scores, name=metric, nbinsx=10, marker_color="#1f77b4"))
fig.update_layout(
title=f"{metric} Score Distribution",
xaxis_title="Score",
yaxis_title="Count",
bargap=0.1,
template="seaborn",
)
figures.append(fig.to_html(full_html=False))
return figures
def create_ci_plot(ci_results: Dict[str, Tuple[float, float, float]]) -> str:
"""Create confidence interval bar plot."""
fig = go.Figure()
for metric, (mean_score, lower, upper) in ci_results.items():
fig.add_trace(
go.Bar(
x=[metric],
y=[mean_score],
error_y=dict(
type="data",
array=[upper - mean_score],
arrayminus=[mean_score - lower],
visible=True,
),
name=metric,
)
)
fig.update_layout(
title="95% confidence interval for all the metrics",
xaxis_title="Metric",
yaxis_title="Score",
template="seaborn",
)
return fig.to_html(full_html=False)
def generate_details_html(metrics_data: List[Dict]) -> List[str]:
"""Generate HTML for detailed metric information."""
details_html = []
metric_details = {}
# Organize metrics by type
for entry in metrics_data:
for metric, values in entry["metrics"].items():
if metric not in metric_details:
metric_details[metric] = []
metric_details[metric].append(
{
"question": entry["question"],
"answer": entry["answer"],
"golden_answer": entry["golden_answer"],
"reason": values.get("reason", ""),
"score": values["score"],
}
)
for metric, details in metric_details.items():
details_html.append(f"{metric} Details
")
details_html.append("""
| Question |
Answer |
Golden Answer |
Reason |
Score |
""")
for item in details:
details_html.append(
f""
f"| {item['question']} | "
f"{item['answer']} | "
f"{item['golden_answer']} | "
f"{item['reason']} | "
f"{item['score']} | "
f"
"
)
details_html.append("
")
return details_html
def get_dashboard_html_template(
figures: List[str], details_html: List[str], benchmark: str = ""
) -> str:
"""Generate the complete HTML dashboard template."""
return f"""
LLM Evaluation Dashboard {benchmark}
LLM Evaluation Metrics Dashboard {benchmark}
Metrics Distribution
{"".join([f'{fig}
' for fig in figures[:-1]])}
95% confidence interval for all the metrics
{figures[-1]}
Detailed Explanations
{"".join(details_html)}
"""
def create_dashboard(
metrics_path: str,
aggregate_metrics_path: str,
output_file: str = "dashboard_with_ci.html",
benchmark: str = "",
) -> str:
"""Create and save the dashboard with all visualizations."""
# Read metrics files
with open(metrics_path, "r") as f:
metrics_data = json.load(f)
with open(aggregate_metrics_path, "r") as f:
aggregate_data = json.load(f)
# Extract data for visualizations
metrics_by_type = defaultdict(list)
for entry in metrics_data:
for metric, values in entry["metrics"].items():
metrics_by_type[metric].append(values["score"])
# Generate visualizations
distribution_figures = create_distribution_plots(metrics_by_type)
ci_plot = create_ci_plot(
{
metric: (data["mean"], data["ci_lower"], data["ci_upper"])
for metric, data in aggregate_data.items()
}
)
# Combine all figures
figures = distribution_figures + [ci_plot]
# Generate HTML components
details_html = generate_details_html(metrics_data)
dashboard_html = get_dashboard_html_template(figures, details_html, benchmark)
# Write to file
with open(output_file, "w", encoding="utf-8") as f:
f.write(dashboard_html)
return output_file