Observability for Hugging Face Datasets with Opik

1
import os
2
import getpass
3
4
if "HUGGINGFACE_HUB_TOKEN" not in os.environ:
5
    os.environ["HUGGINGFACE_HUB_TOKEN"] = getpass.getpass("Enter your Hugging Face token: ")
6
7
# Set project name for organization
8
os.environ["OPIK_PROJECT_NAME"] = "huggingface-datasets-integration-demo"

1
from datasets import load_dataset, Dataset as HFDataset
2
from opik import Opik
3
from typing import Optional, Dict, Any, List
4
import json
5
from tqdm import tqdm
6
import warnings
7
import numpy as np
8
import pandas as pd
9
10
warnings.filterwarnings('ignore')
11
12
class HuggingFaceToOpikConverter:
13
    """Utility class to convert Hugging Face datasets to Opik format."""
14
15
    def __init__(self, opik_client: Opik):
16
        self.opik_client = opik_client
17
18
    def load_hf_dataset(
19
        self,
20
        dataset_name: str,
21
        split: Optional[str] = None,
22
        config: Optional[str] = None,
23
        subset_size: Optional[int] = None,
24
        **kwargs
25
    ) -> HFDataset:
26
        """
27
        Load a dataset from Hugging Face Hub.
28
29
        Args:
30
            dataset_name: Name of the dataset on HF Hub
31
            split: Specific split to load (train, validation, test)
32
            config: Configuration/subset of the dataset
33
            subset_size: Limit the number of samples
34
            **kwargs: Additional arguments for load_dataset
35
36
        Returns:
37
            Loaded Hugging Face dataset
38
        """
39
        print(f"📥 Loading dataset: {dataset_name}")
40
        if config:
41
            print(f"   Config: {config}")
42
        if split:
43
            print(f"   Split: {split}")
44
45
        # Load the dataset
46
        dataset = load_dataset(
47
            dataset_name,
48
            name=config,
49
            split=split,
50
            **kwargs
51
        )
52
53
        # Limit dataset size if specified
54
        if subset_size and len(dataset) > subset_size:
55
            dataset = dataset.select(range(subset_size))
56
            print(f"   Limited to {subset_size} samples")
57
58
        print(f"   ✅ Loaded {len(dataset)} samples")
59
        print(f"   Features: {list(dataset.features.keys())}")
60
61
        return dataset

1
# Initialize the converter
2
opik_client = Opik()
3
converter = HuggingFaceToOpikConverter(opik_client)
4
5
# Load a dataset from Hugging Face
6
dataset = converter.load_hf_dataset(
7
    dataset_name="squad",
8
    split="validation",
9
    subset_size=100  # Limit for demo
10
)
11
12
# Convert to Opik format
13
opik_data = converter.convert_to_opik_format(
14
    dataset=dataset,
15
    input_columns=["question"],
16
    output_columns=["answers"],
17
    metadata_columns=["id", "title"],
18
    dataset_name="squad-qa-dataset",
19
    description="SQuAD question answering dataset converted from Hugging Face"
20
)
21
22
print(f"✅ Converted {len(opik_data)} items to Opik format!")

1
def convert_to_opik_format(
2
    self,
3
    dataset: HFDataset,
4
    input_columns: List[str],
5
    output_columns: List[str],
6
    metadata_columns: Optional[List[str]] = None,
7
    dataset_name: str = "huggingface-dataset",
8
    description: str = "Dataset converted from Hugging Face"
9
) -> List[Dict[str, Any]]:
10
    """
11
    Convert a Hugging Face dataset to Opik format.
12
13
    Args:
14
        dataset: Hugging Face dataset
15
        input_columns: List of column names to use as input
16
        output_columns: List of column names to use as expected output
17
        metadata_columns: Optional list of columns to include as metadata
18
        dataset_name: Name for the Opik dataset
19
        description: Description for the Opik dataset
20
21
    Returns:
22
        List of Opik dataset items
23
    """
24
    opik_items = []
25
26
    for row in tqdm(dataset, desc="Converting to Opik format"):
27
        # Extract input data
28
        input_data = {}
29
        for col in input_columns:
30
            if col in dataset.features:
31
                input_data[col] = self._extract_field_value(row, col)
32
33
        # Extract expected output
34
        expected_output = {}
35
        for col in output_columns:
36
            if col in dataset.features:
37
                expected_output[col] = self._extract_field_value(row, col)
38
39
        # Extract metadata
40
        metadata = {}
41
        if metadata_columns:
42
            for col in metadata_columns:
43
                if col in dataset.features:
44
                    metadata[col] = self._extract_field_value(row, col)
45
46
        # Create Opik dataset item
47
        item = {
48
            "input": input_data,
49
            "expected_output": expected_output,
50
            "metadata": metadata
51
        }
52
        opik_items.append(item)
53
54
    return opik_items

1
from opik import track
2
3
@track
4
def evaluate_qa_model(dataset_item):
5
    """Evaluate a Q&A model using Hugging Face dataset."""
6
    question = dataset_item["input"]["question"]
7
8
    # Your model logic here (replace with actual model)
9
    if 'what' in question.lower():
10
        response = "This is a question asking for information."
11
    elif 'how' in question.lower():
12
        response = "This is a question asking for a process or method."
13
    else:
14
        response = "This is a general question that requires analysis."
15
16
    return {
17
        "question": question,
18
        "response": response,
19
        "expected": dataset_item["expected_output"],
20
        "metadata": dataset_item["metadata"]
21
    }
22
23
# Evaluate on your dataset
24
for item in opik_data[:5]:  # Evaluate first 5 items
25
    result = evaluate_qa_model(item)
26
    print(f"Question: {result['question'][:50]}...")

1
# Load SQuAD dataset
2
squad_dataset = converter.load_hf_dataset(
3
    dataset_name="squad",
4
    split="validation",
5
    subset_size=50
6
)
7
8
# Convert to Opik format
9
squad_opik = converter.convert_to_opik_format(
10
    dataset=squad_dataset,
11
    input_columns=["question"],
12
    output_columns=["answers"],
13
    metadata_columns=["id", "title"],
14
    dataset_name="squad-qa-dataset",
15
    description="SQuAD question answering dataset"
16
)

1
# Load GLUE SST-2 dataset
2
sst2_dataset = converter.load_hf_dataset(
3
    dataset_name="glue",
4
    config_name="sst2",
5
    split="validation",
6
    subset_size=100
7
)
8
9
# Convert to Opik format
10
sst2_opik = converter.convert_to_opik_format(
11
    dataset=sst2_dataset,
12
    input_columns=["sentence"],
13
    output_columns=["label"],
14
    metadata_columns=["idx"],
15
    dataset_name="sst2-sentiment-dataset",
16
    description="SST-2 sentiment analysis dataset from GLUE"
17
)

1
# Load Common Crawl dataset
2
cc_dataset = converter.load_hf_dataset(
3
    dataset_name="common_crawl",
4
    subset_size=200
5
)
6
7
# Convert to Opik format
8
cc_opik = converter.convert_to_opik_format(
9
    dataset=cc_dataset,
10
    input_columns=["text"],
11
    output_columns=["language"],
12
    metadata_columns=["url", "timestamp"],
13
    dataset_name="common-crawl-dataset",
14
    description="Common Crawl text classification dataset"
15
)

1
from opik.evaluation import evaluate
2
from opik.evaluation.metrics import Hallucination
3
4
# Define your evaluation task
5
def evaluation_task(x):
6
    return {
7
        "message": x["input"]["question"],
8
        "output": x["response"],
9
        "reference": x["expected_output"]["answers"]
10
    }
11
12
# Create the Hallucination metric
13
hallucination_metric = Hallucination()
14
15
# Run the evaluation
16
evaluation_results = evaluate(
17
    experiment_name="huggingface-datasets-evaluation",
18
    dataset=squad_opik,
19
    task=evaluation_task,
20
    scoring_metrics=[hallucination_metric],
21
    project_name="my-project",
22
)