Alexander Watson
analysis improvements
0629e69
raw
history blame
22.9 kB
import base64
import datetime
import io
import json
import re
from collections import Counter
import pandas as pd
import plotly.express as px
import plotly.graph_objects as go
import tiktoken
import yaml
from openai import OpenAI
def extract_json_from_response(text: str) -> str:
"""Extract JSON from a response that might contain markdown code blocks."""
# Try to find JSON within code blocks first
json_match = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", text, re.DOTALL)
if json_match:
return json_match.group(1)
# If no code blocks, try to find raw JSON
json_match = re.search(r"\{.*\}", text, re.DOTALL)
if json_match:
return json_match.group(0)
# If no JSON found, return the original text
return text
def count_tokens(text: str, model: str = "gpt-4") -> int:
"""Count tokens in text using tiktoken."""
try:
encoder = tiktoken.encoding_for_model(model)
return len(encoder.encode(str(text)))
except Exception as e:
print(f"Error counting tokens: {e}")
return 0
def create_distribution_plot(data, column):
"""Create a distribution plot using Plotly and convert to image."""
try:
# Check if the column contains lists
if isinstance(data[column].iloc[0], list):
print(f"Processing list column: {column}")
value_counts = flatten_list_column(data, column)
fig = go.Figure(
[
go.Bar(
x=value_counts.index,
y=value_counts.values,
marker=dict(
color=value_counts.values,
colorscale=px.colors.sequential.Plotly3,
),
)
]
)
else:
if data[column].dtype in ["int64", "float64"]:
# Continuous data - use histogram
fig = go.Figure()
fig.add_trace(
go.Histogram(
x=data[column],
name="Count",
nbinsx=30,
marker=dict(
color="rgba(110, 68, 255, 0.7)",
line=dict(color="rgba(184, 146, 255, 1)", width=1),
),
)
)
else:
# Categorical data
value_counts = data[column].value_counts()
fig = go.Figure(
[
go.Bar(
x=value_counts.index,
y=value_counts.values,
marker=dict(
color=value_counts.values,
colorscale=px.colors.sequential.Plotly3,
),
)
]
)
# Common layout updates
fig.update_layout(
title=dict(text=f"Distribution of {column}", x=0.5, y=0.95),
xaxis_title=column,
yaxis_title="Count",
template="plotly_white",
margin=dict(t=50, l=50, r=30, b=50),
width=600,
height=400,
showlegend=False,
plot_bgcolor="rgba(0,0,0,0)",
paper_bgcolor="rgba(0,0,0,0)",
)
# Rotate x-axis labels if needed
if isinstance(data[column].iloc[0], list) or data[column].dtype not in [
"int64",
"float64",
]:
fig.update_layout(xaxis_tickangle=-45)
# Update grid style
fig.update_yaxes(gridcolor="rgba(128,128,128,0.1)", gridwidth=1)
fig.update_xaxes(gridcolor="rgba(128,128,128,0.1)", gridwidth=1)
# Convert to PNG with moderate resolution
img_bytes = fig.to_image(format="png", scale=1.5)
# Encode to base64
img_base64 = base64.b64encode(img_bytes).decode()
return img_base64
except Exception as e:
print(f"Error creating distribution plot for {column}: {str(e)}")
raise e
def create_wordcloud(data, column):
"""Create a word cloud visualization."""
import matplotlib.pyplot as plt
from wordcloud import WordCloud
try:
# Handle list columns
if isinstance(data[column].iloc[0], list):
text = " ".join(
[
" ".join(map(str, sublist))
for sublist in data[column]
if isinstance(sublist, list)
]
)
else:
# Handle regular columns
text = " ".join(data[column].astype(str))
wordcloud = WordCloud(
width=600,
height=300,
background_color="white",
colormap="plasma",
max_words=100,
).generate(text)
# Create matplotlib figure
plt.figure(figsize=(8, 4))
plt.imshow(wordcloud, interpolation="bilinear")
plt.axis("off")
plt.title(f"Word Cloud for {column}")
# Save to bytes
buf = io.BytesIO()
plt.savefig(buf, format="png", bbox_inches="tight", dpi=150)
plt.close()
buf.seek(0)
# Convert to base64
img_base64 = base64.b64encode(buf.getvalue()).decode()
return img_base64
except Exception as e:
print(f"Error creating word cloud for {column}: {str(e)}")
raise e
def analyze_dataset_with_openai(client: OpenAI, data) -> dict:
"""Analyze dataset using OpenAI API with improved type inference and efficient sampling."""
# Convert dictionary to DataFrame if needed
if isinstance(data, dict):
df = pd.DataFrame(data)
else:
df = data
# Take a very small sample for efficiency
sample_size = min(3, len(df))
if len(df) > 3:
sample_indices = df.index[
:sample_size
] # Take first 3 rows instead of random sampling
sample_df = df.loc[sample_indices]
else:
sample_df = df
dataset_sample = sample_df.to_dict("records")
single_record = dataset_sample[0]
# Create type hints dictionary - only process the sample
type_hints = {}
for column in sample_df.columns:
# Get the pandas dtype
dtype = sample_df[column].dtype
# Efficiently identify types without complex operations
if pd.api.types.is_integer_dtype(dtype):
type_hints[column] = "integer"
elif pd.api.types.is_float_dtype(dtype):
type_hints[column] = "number"
elif pd.api.types.is_bool_dtype(dtype):
type_hints[column] = "boolean"
elif pd.api.types.is_datetime64_any_dtype(dtype):
type_hints[column] = "datetime"
elif pd.api.types.is_categorical_dtype(dtype):
type_hints[column] = "categorical"
elif pd.api.types.is_string_dtype(dtype):
# Simple check for list-like values
first_val = sample_df[column].iloc[0]
if isinstance(first_val, list):
type_hints[column] = "array"
else:
type_hints[column] = "string"
else:
type_hints[column] = "unknown"
prompt = f"""Analyze this dataset sample and provide the following in a JSON response:
1. A concise description that includes:
- A one-sentence overview of what the dataset contains
- A bullet-pointed list of key features and statistics
- A brief statement about potential ML/AI applications
2. A schema showing each field's type and description. Here is the actual DataFrame type information:
{json.dumps(type_hints, indent=2)}
And here's a single record for reference:
{json.dumps(single_record, indent=2)}
3. A formatted example record
Format your response as JSON with these exact keys:
{{
"description": {{
"overview": "One clear sentence describing the dataset...",
"key_features": [
"Feature or statistic 1",
"Feature or statistic 2"
],
"ml_applications": "Brief statement about ML/AI use cases..."
}},
"schema": {{
"field_name": {{
"type": "use the type from the provided type_hints",
"description": "Description of what this field contains"
}}
}},
"example": {{"key": "value"}}
}}
For context, here are more sample records:
{json.dumps(dataset_sample, indent=2)}
"""
try:
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": prompt}],
temperature=0.7,
max_tokens=2000,
)
# Get the response content
response_text = response.choices[0].message.content
# Extract JSON from the response
json_str = extract_json_from_response(response_text)
# Parse the JSON
result = json.loads(json_str)
return result
except Exception as e:
print(f"OpenAI API error: {str(e)}")
return {
"description": {
"overview": "Error analyzing dataset",
"key_features": ["Error: Failed to analyze dataset"],
"ml_applications": "Analysis unavailable",
},
"schema": {},
"example": {},
}
def analyze_dataset_statistics(df):
"""Generate simplified dataset statistics with token counting."""
stats = {
"basic_stats": {
"total_records": len(df),
"total_features": len(df.columns),
"memory_usage": f"{df.memory_usage(deep=True).sum() / (1024*1024):.2f} MB",
},
"token_stats": {"total": 0, "by_column": {}},
}
# Count tokens for each column
for column in df.columns:
try:
if df[column].dtype == "object" or isinstance(df[column].iloc[0], list):
# For list columns, join items into strings
if isinstance(df[column].iloc[0], list):
token_counts = df[column].apply(
lambda x: count_tokens(" ".join(str(item) for item in x))
)
else:
token_counts = df[column].apply(lambda x: count_tokens(str(x)))
total_tokens = int(token_counts.sum())
stats["token_stats"]["total"] += total_tokens
stats["token_stats"]["by_column"][column] = total_tokens
except Exception as e:
print(f"Error processing column {column}: {str(e)}")
continue
return stats
def format_dataset_stats(stats):
"""Format simplified dataset statistics as markdown."""
md = """## Dataset Overview
### Basic Statistics
* Total Records: {total_records:,}
* Total Features: {total_features}
* Memory Usage: {memory_usage}
""".format(
**stats["basic_stats"]
)
# Token Statistics
if stats["token_stats"]["total"] > 0:
md += "\n### Token Info\n"
md += f"* Total Tokens: {stats['token_stats']['total']:,}\n"
if stats["token_stats"]["by_column"]:
md += "\nTokens by Column:\n"
for col, count in stats["token_stats"]["by_column"].items():
md += f"* {col}: {count:,}\n"
return md
def generate_dataset_card(
dataset_info: dict,
distribution_plots: dict,
wordcloud_plots: dict,
openai_analysis: dict,
df: pd.DataFrame,
) -> str:
"""Generate a beautiful and clean dataset card."""
# Basic dataset metadata
yaml_content = {
"language": ["en"],
"license": "apache-2.0",
"multilinguality": "monolingual",
"size_categories": [get_size_category(len(df))],
"task_categories": ["other"],
}
yaml_string = yaml.dump(yaml_content, sort_keys=False)
# Generate dataset statistics
stats = analyze_dataset_statistics(df)
description = openai_analysis["description"]
# Build the markdown content with proper spacing
readme_content = f"""---
{yaml_string}---
# {dataset_info['dataset_name']}
{description['overview']}
### Key Features
{chr(10).join(f'* {feature}' for feature in description['key_features'])}
### Potential Applications
{description['ml_applications']}
## Dataset Statistics
* Total Records: {stats['basic_stats']['total_records']:,}
* Total Features: {stats['basic_stats']['total_features']}
* Memory Usage: {stats['basic_stats']['memory_usage']}
## Dataset Schema
| Field | Type | Description |
| --- | --- | --- |
{chr(10).join(f"| {field} | {info['type']} | {info['description']} |" for field, info in openai_analysis['schema'].items())}
## Example Record
```json
{json.dumps(openai_analysis['example'], indent=2)}
```
## Data Distribution Analysis
The following visualizations show the distribution patterns and characteristics of key features in the dataset:
"""
# Add individual distribution plots with clean spacing
for col, img_str in distribution_plots.items():
readme_content += f"""### Distribution of {col}
<img src="data:image/png;base64,{img_str}" alt="Distribution of {col}" style="max-width: 800px;">
"""
# Add word clouds with clean spacing
if wordcloud_plots:
readme_content += "## Feature Word Clouds\n\n"
for col, img_str in wordcloud_plots.items():
readme_content += f"""### Word Cloud for {col}
<img src="data:image/png;base64,{img_str}" alt="Word Cloud for {col}" style="max-width: 800px;">
"""
# Add token statistics if available
if stats.get("token_stats") and stats["token_stats"]["total"] > 0:
readme_content += """## Token Statistics
"""
readme_content += f"* Total Tokens: {stats['token_stats']['total']:,}\n"
if stats["token_stats"].get("by_column"):
readme_content += "\n**Tokens by Column:**\n"
for col, count in stats["token_stats"]["by_column"].items():
readme_content += f"* {col}: {count:,}\n"
# Add citation section
clean_name = dataset_info["dataset_name"].replace("/", "_")
readme_content += f"""
## Citation
```bibtex
@dataset{{{clean_name},
title = {{{dataset_info['dataset_name']}}},
year = {{{datetime.datetime.now().year}}},
publisher = {{Hugging Face}},
url = {{https://huggingface.co/datasets/{dataset_info['dataset_name']}}}
}}
```
### Usage Guidelines
This dataset is released under the Apache 2.0 License. When using this dataset:
* πŸ“š Cite the dataset using the BibTeX entry above
* 🀝 Consider contributing improvements or reporting issues
* πŸ’‘ Share derivative works with the community when possible
"""
return readme_content
def get_size_category(record_count: int) -> str:
"""Determine the size category based on record count."""
if record_count < 1000:
return "n<1K"
elif record_count < 10000:
return "1K<n<10K"
elif record_count < 100000:
return "10K<n<100K"
elif record_count < 1000000:
return "100K<n<1M"
else:
return "n>1M"
def format_overview_section(analysis: dict, stats: dict) -> str:
"""Create a comprehensive overview section."""
description = analysis["description"]
overview = f"""
{description['overview']}
### Key Features and Characteristics
{chr(10).join(f'* {feature}' for feature in description['key_features'])}
### Potential Applications
{description['ml_applications']}
### Dataset Size
* Total Records: {stats['basic_stats']['total_records']:,}
* Total Features: {stats['basic_stats']['total_features']}
* Memory Usage: {stats['basic_stats']['memory_usage']}
"""
return overview.strip()
def format_schema_section(schema: dict, df: pd.DataFrame) -> str:
"""Generate an enhanced schema section with statistics."""
# Table header
table = "| Field | Type | Description | Non-Null Count | Unique Values |\n"
table += "| --- | --- | --- | --- | --- |\n"
# Generate rows with additional statistics
for field, info in schema.items():
try:
non_null = df[field].count()
unique = df[field].nunique()
row = f"| {field} | {info['type']} | {info['description']} | {non_null:,} | {unique:,} |"
table += row + "\n"
except Exception as e:
print(f"Error processing field {field}: {e}")
continue
return table
def format_visualization_section(
distribution_plots: dict, wordcloud_plots: dict
) -> str:
"""Format the visualization section with improved layout."""
content = (
"""The following visualizations show key characteristics of the dataset:\n\n"""
)
# Add distribution plots
if distribution_plots:
content += "### Distribution Plots\n\n"
content += '<div style="display: grid; grid-template-columns: repeat(auto-fit, minmax(500px, 1fr)); gap: 20px;">\n'
for col, img_str in distribution_plots.items():
content += f"""<div>
<h4>Distribution of {col}</h4>
<img src="data:image/png;base64,{img_str}" style="width: 100%; height: auto;">
</div>\n"""
content += "</div>\n\n"
# Add word clouds
if wordcloud_plots:
content += "### Word Clouds\n\n"
content += '<div style="display: grid; grid-template-columns: repeat(auto-fit, minmax(400px, 1fr)); gap: 20px;">\n'
for col, img_str in wordcloud_plots.items():
content += f"""<div>
<h4>Word Cloud for {col}</h4>
<img src="data:image/png;base64,{img_str}" style="width: 100%; height: auto;">
</div>\n"""
content += "</div>\n"
return content
def generate_limitations_section(df: pd.DataFrame, analysis: dict) -> str:
"""Generate a section about dataset limitations and potential biases."""
limitations = [
"This dataset may not be representative of all possible scenarios or use cases.",
f"The dataset contains {len(df):,} records, which may limit its applicability to certain tasks.",
"There may be inherent biases in the data collection or annotation process.",
]
# Add warnings about missing values if present
missing_values = df.isnull().sum()
if missing_values.any():
limitations.append(
f"Some fields contain missing values: {', '.join(missing_values[missing_values > 0].index)}"
)
return f"""The following limitations and potential biases should be considered when using this dataset:
{chr(10).join(f'* {limitation}' for limitation in limitations)}
Please consider these limitations when using the dataset and validate results accordingly."""
def generate_usage_section(dataset_info: dict, analysis: dict) -> str:
"""Generate comprehensive usage guidelines."""
return f"""This dataset is released under the Apache 2.0 License. When using this dataset:
* πŸ“š Cite the dataset using the BibTeX entry provided below
* 🀝 Consider contributing improvements or reporting issues
* πŸ’‘ Share derivative works with the community when possible
* πŸ” Validate the dataset's suitability for your specific use case
* ⚠️ Be aware of the limitations and biases discussed above
* πŸ“Š Consider the dataset size and computational requirements for your application
For questions or additional information, please visit the dataset repository on Hugging Face.
"""
def get_task_categories(df: pd.DataFrame, analysis: dict) -> list:
"""Infer potential task categories based on the data and analysis."""
categories = ["other"] # Default category
# Add more sophisticated task inference logic based on column names and content
text_columns = df.select_dtypes(include=["object"]).columns
numeric_columns = df.select_dtypes(include=["int64", "float64"]).columns
if len(text_columns) > 0:
categories.append("text-classification")
if len(numeric_columns) > 0:
categories.append("regression")
return list(set(categories)) # Remove duplicates
def clean_dataset_name(name: str) -> str:
"""Clean dataset name for citation."""
return name.replace("/", "_").replace("-", "_").lower()
def generate_schema_table(schema: dict) -> str:
"""Generate a markdown table for the schema, handling nested structures."""
# Table header
table = "| Field | Type | Description |\n| --- | --- | --- |\n"
# Generate rows recursively
rows = []
for field, info in schema.items():
rows.extend(format_schema_item(field, info))
# Join all rows
table += "\n".join(rows)
return table
def format_stats_section(stats: dict) -> str:
"""Format the statistics section of the dataset card."""
content = """### Basic Statistics
"""
# Add basic stats
for key, value in stats["basic_stats"].items():
# Convert key from snake_case to Title Case
formatted_key = key.replace("_", " ").title()
content += f"* {formatted_key}: {value}\n"
# Add token statistics if available
if stats.get("token_stats") and stats["token_stats"]["total"] > 0:
content += "\n### Token Statistics\n"
content += f"* Total Tokens: {stats['token_stats']['total']:,}\n"
if stats["token_stats"].get("by_column"):
content += "\n**Tokens by Column:**\n"
for col, count in stats["token_stats"]["by_column"].items():
content += f"* {col}: {count:,}\n"
return content
def format_schema_item(field_name: str, field_info: dict, prefix: str = "") -> list:
"""Recursively format schema items for nested structures."""
rows = []
# Handle nested objects
if isinstance(field_info, dict):
if "type" in field_info and "description" in field_info:
# This is a leaf node with type and description
rows.append(
f"| {prefix}{field_name} | {field_info['type']} | {field_info['description']} |"
)
else:
# This is a nested object, recurse through its properties
for subfield, subinfo in field_info.items():
if prefix:
new_prefix = f"{prefix}{field_name}."
else:
new_prefix = f"{field_name}."
rows.extend(format_schema_item(subfield, subinfo, new_prefix))
return rows
def flatten_list_column(data, column):
"""Flatten a column containing lists into individual values with counts."""
# Flatten the lists into individual items
flattened = [
item
for sublist in data[column]
if isinstance(sublist, list)
for item in sublist
]
# Count occurrences
value_counts = pd.Series(Counter(flattened))
return value_counts