Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/Arize-ai/phoenix/llms.txt

Use this file to discover all available pages before exploring further.

Phoenix automatically versions your datasets, ensuring reproducibility and enabling you to track changes over time. Every modification to a dataset creates a new version, while experiments remain linked to their original versions.

​
Understanding Versions

Dataset versions are created automatically when you:
  • Create a new dataset - Creates version 1
  • Add examples - Creates a new version with added examples
  • Modify examples - Creates a new version with changes (future feature)
Each version has:
  • A unique version_id
  • A created_at timestamp
  • Optional description and metadata
  • Links to all experiments run on that version
from phoenix.client import Client

client = Client()

# Create initial dataset (version 1)
dataset_v1 = client.datasets.create_dataset(
    name="qa-dataset",
    inputs=[{"question": "What is AI?"}],
    outputs=[{"answer": "Artificial Intelligence is..."}]
)

print(f"Version ID: {dataset_v1.version_id}")
print(f"Created at: {dataset_v1.created_at}")

# Add more examples (creates version 2)
dataset_v2 = client.datasets.add_examples_to_dataset(
    dataset="qa-dataset",
    inputs=[{"question": "What is ML?"}],
    outputs=[{"answer": "Machine Learning is..."}]
)

print(f"New version ID: {dataset_v2.version_id}")
print(f"Example count v1: {len(dataset_v1)}")
print(f"Example count v2: {len(dataset_v2)}")

​
Listing Versions

Retrieve all versions of a dataset: 
from phoenix.client import Client

client = Client()

# Get all versions (latest first)
versions = client.datasets.get_dataset_versions(
    dataset="qa-dataset",
    limit=100  # Get up to 100 most recent versions
)

# Display version history
for version in versions:
    print(f"Version: {version['version_id']}")
    print(f"  Created: {version['created_at']}")
    if version.get('description'):
        print(f"  Description: {version['description']}")
    if version.get('metadata'):
        print(f"  Metadata: {version['metadata']}")
    print()

​
Retrieving Specific Versions

Access any historical version by its version_id: 
from phoenix.client import Client

client = Client()

# Get latest version (default)
latest = client.datasets.get_dataset(dataset="qa-dataset")
print(f"Latest version: {latest.version_id} with {len(latest)} examples")

# Get specific version
specific = client.datasets.get_dataset(
    dataset="qa-dataset",
    version_id="v_abc123def456"
)
print(f"Specific version: {specific.version_id} with {len(specific)} examples")

# Compare versions
print(f"Examples added: {len(latest) - len(specific)}")

​
Version Reproducibility

Experiments are permanently linked to specific dataset versions: 
from phoenix.client import Client
from phoenix.experiments import run_experiment

client = Client()

# Get dataset at current version
dataset_v1 = client.datasets.get_dataset(dataset="qa-dataset")

# Run experiment on v1
experiment_v1 = run_experiment(
    dataset=dataset_v1,
    task=my_task,
    experiment_name="test-on-v1"
)

print(f"Experiment dataset version: {dataset_v1.version_id}")

# Add more examples (creates v2)
dataset_v2 = client.datasets.add_examples_to_dataset(
    dataset="qa-dataset",
    inputs=[{"question": "New question"}],
    outputs=[{"answer": "New answer"}]
)

# Run experiment on v2
experiment_v2 = run_experiment(
    dataset=dataset_v2,
    task=my_task,
    experiment_name="test-on-v2"
)

print(f"Experiment v1 still references: {dataset_v1.version_id}")
print(f"Experiment v2 references: {dataset_v2.version_id}")

# You can re-run on old versions
rerun_v1 = run_experiment(
    dataset=client.datasets.get_dataset(
        dataset="qa-dataset",
        version_id=dataset_v1.version_id
    ),
    task=improved_task,
    experiment_name="improved-on-v1"
)

​
Exporting Datasets

​
Export to DataFrame

Convert datasets to pandas DataFrames for analysis or storage: 
from phoenix.client import Client
import pandas as pd

client = Client()

dataset = client.datasets.get_dataset(dataset="qa-dataset")

# Export to DataFrame
df = dataset.to_dataframe()
print(df.head())
print(df.columns)
# Index(['input', 'output', 'metadata'], dtype='object')
# Index name: 'example_id'

# Save to CSV
df.to_csv("dataset_export.csv")

# Save to Parquet
df.to_parquet("dataset_export.parquet")

​
Export to JSON

Export datasets in a portable JSON format: 
import json
from phoenix.client import Client

client = Client()

dataset = client.datasets.get_dataset(dataset="qa-dataset")

# Convert to dict
dataset_dict = dataset.to_dict()

# Save to JSON file
with open("dataset_export.json", "w") as f:
    json.dump(dataset_dict, f, indent=2, default=str)

print("Exported dataset to JSON")

​
Import from JSON

Restore datasets from JSON exports: 
import json
from phoenix.client import Client
from phoenix.client.resources.datasets import Dataset

client = Client()

# Load from JSON
with open("dataset_export.json", "r") as f:
    dataset_dict = json.load(f)

# Restore Dataset object
restored = Dataset.from_dict(dataset_dict)

print(f"Restored dataset: {restored.name}")
print(f"Version: {restored.version_id}")
print(f"Examples: {len(restored)}")

# Can now use in experiments
from phoenix.experiments import run_experiment

experiment = run_experiment(
    dataset=restored,
    task=my_task,
    experiment_name="test-restored"
)

​
Export Specific Versions

Export any version for archival or sharing: 
from phoenix.client import Client
import pandas as pd

client = Client()

# Get all versions
versions = client.datasets.get_dataset_versions(dataset="qa-dataset")

# Export each version
for version in versions:
    dataset = client.datasets.get_dataset(
        dataset="qa-dataset",
        version_id=version['version_id']
    )
    
    # Save with version in filename
    filename = f"dataset_{version['version_id']}.csv"
    df = dataset.to_dataframe()
    df.to_csv(filename)
    
    print(f"Exported {filename}: {len(dataset)} examples")

​
Export with Splits

Export specific dataset splits: 
from phoenix.client import Client

client = Client()

# Export train split
train_data = client.datasets.get_dataset(
    dataset="ml-dataset",
    splits=["train"]
)
train_df = train_data.to_dataframe()
train_df.to_csv("train_split.csv")

# Export validation split
val_data = client.datasets.get_dataset(
    dataset="ml-dataset",
    splits=["validation"]
)
val_df = val_data.to_dataframe()
val_df.to_csv("validation_split.csv")

# Export test split
test_data = client.datasets.get_dataset(
    dataset="ml-dataset",
    splits=["test"]
)
test_df = test_data.to_dataframe()
test_df.to_csv("test_split.csv")

​
Working with Version Metadata

While Phoenix automatically creates versions, you can track additional context: 
from phoenix.client import Client

client = Client()

# Create dataset with metadata
dataset = client.datasets.create_dataset(
    name="versioned-dataset",
    inputs=[{"question": "Q1"}],
    outputs=[{"answer": "A1"}],
    dataset_description="Initial version with baseline examples"
)

print(f"Version 1: {dataset.version_id}")
print(f"Description: {dataset.description}")

# Add examples with context
updated = client.datasets.add_examples_to_dataset(
    dataset="versioned-dataset",
    inputs=[{"question": "Q2"}, {"question": "Q3"}],
    outputs=[{"answer": "A2"}, {"answer": "A3"}]
)

print(f"Version 2: {updated.version_id}")

# Track version history in your own metadata
versions_log = []
for version in client.datasets.get_dataset_versions(dataset="versioned-dataset"):
    versions_log.append({
        "version_id": version['version_id'],
        "created_at": version['created_at'],
        "description": version.get('description', 'N/A')
    })

print("Version history:")
for log in versions_log:
    print(f"  {log['created_at']}: {log['version_id']} - {log['description']}")

​
Version Cleanup Strategy

While Phoenix keeps all versions, you can implement your own archival strategy: 
from datetime import datetime, timedelta, timezone
from phoenix.client import Client

client = Client()

# Find versions older than 90 days
cutoff = datetime.now(timezone.utc) - timedelta(days=90)
versions = client.datasets.get_dataset_versions(dataset="qa-dataset")

old_versions = [
    v for v in versions
    if v['created_at'] < cutoff
]

print(f"Found {len(old_versions)} versions older than 90 days")

​
Best Practices

Pin Versions

Reference specific version_ids in production code for reproducibility.

Document Changes

Use descriptive dataset_description when creating versions to track what changed.

Regular Exports

Periodically export important versions to external storage for backup.

Version Comparison

Compare experiments across versions to understand impact of dataset changes.

​
Next Steps

Creating Datasets

Learn how to create and populate datasets

Running Experiments

Run experiments on your versioned datasets

Build docs developers (and LLMs) love

Get started for freeTalk to us