Dataset & Batch

Fair Forge uses two primary data structures to represent conversation data: Dataset and Batch.

Dataset

A Dataset represents a complete conversation session with an AI assistant.

from fair_forge.schemas.common import Dataset

dataset = Dataset(
    session_id="session-123",
    assistant_id="my-assistant-v1",
    language="english",
    context="You are a helpful customer service assistant.",
    conversation=[...],  # list[Batch]
)

Fields

Field	Type	Required	Description
`session_id`	`str`	Yes	Unique identifier for the conversation session
`assistant_id`	`str`	Yes	Identifier of the assistant being evaluated
`language`	`str \| None`	No	Language of the conversation (e.g., “english”, “spanish”)
`context`	`str`	Yes	System context/instructions provided to the assistant
`conversation`	`list[Batch]`	Yes	List of Q&A interactions in the conversation

Example

from fair_forge.schemas.common import Dataset, Batch

dataset = Dataset(
    session_id="customer-support-001",
    assistant_id="support-bot-v2",
    language="english",
    context="""You are a helpful customer service assistant for TechStore.
    Be polite, concise, and always offer to help further.""",
    conversation=[
        Batch(
            qa_id="q1",
            query="Hi, I need help with my order",
            assistant="Hello! I'd be happy to help with your order. Could you please provide your order number?",
            ground_truth_assistant="Greet the customer and ask for order number.",
        ),
        Batch(
            qa_id="q2",
            query="It's ORDER-12345",
            assistant="Thank you! I found your order. It was shipped yesterday and should arrive by Friday. Is there anything else I can help with?",
            ground_truth_assistant="Look up the order and provide shipping status.",
        ),
    ],
)

Batch

A Batch represents a single question-answer interaction within a conversation.

from fair_forge.schemas.common import Batch

batch = Batch(
    qa_id="q1",
    query="What is the capital of France?",
    assistant="The capital of France is Paris.",
    ground_truth_assistant="Paris is the capital of France.",
)

Fields

Field	Type	Required	Description
`qa_id`	`str`	Yes	Unique identifier for this interaction
`query`	`str`	Yes	User’s question or message
`assistant`	`str`	Yes	Assistant’s response
`ground_truth_assistant`	`str \| None`	No	Expected/ideal response for comparison
`observation`	`str \| None`	No	Additional notes or observations
`agentic`	`dict \| None`	No	Metadata about the interaction
`ground_truth_agentic`	`dict \| None`	No	Expected metadata
`logprobs`	`dict \| None`	No	Log probabilities from the model
`weight`	`float \| None`	No	Relative importance when aggregating session-level scores

Field Details

qa_id

A unique identifier for the interaction within the conversation. Use a consistent naming scheme:

qa_id="q1"           # Simple numbering
qa_id="order-inquiry-001"  # Descriptive naming
qa_id="session123_turn5"   # Session-based

query

The user’s input message or question. This is what the assistant is responding to:

query="What are your business hours?"
query="Can you help me debug this Python code?"
query="Tell me a joke about programming"

assistant

The assistant’s actual response. This is what gets evaluated:

assistant="Our business hours are Monday to Friday, 9 AM to 5 PM EST."

ground_truth_assistant

The expected or ideal response. Used by some metrics for comparison:

ground_truth_assistant="Mon-Fri 9-5 EST"

This is optional but recommended for metrics like Context and Conversational.

observation

Additional context or notes about the interaction:

observation="Customer seems frustrated in this exchange"
observation="This is a follow-up to the previous question"

agentic

Metadata dictionary for storing additional information:

agentic={
    "difficulty": "medium",
    "query_type": "factual",
    "chunk_id": "doc_section_3",
    "turn_number": 2,
}

Used by generators to store query metadata.

logprobs

Log probabilities from the model (if available):

logprobs={
    "tokens": ["The", "capital", "is", "Paris"],
    "token_logprobs": [-0.1, -0.05, -0.02, -0.01],
}

weight

Optional relative importance of this interaction when metrics aggregate scores at session level (Conversational, Context, Regulatory). Controls how much each QA pair contributes to the session mean.

# All provided and sum to 1 → used as-is
Batch(qa_id="q1", ..., weight=0.5),
Batch(qa_id="q2", ..., weight=0.3),
Batch(qa_id="q3", ..., weight=0.2),

# Some provided → remaining weight split equally among unweighted
Batch(qa_id="q1", ..., weight=0.6),
Batch(qa_id="q2", ...),   # gets 0.2
Batch(qa_id="q3", ...),   # gets 0.2

# None provided → equal weights (1/n each)
Batch(qa_id="q1", ...),
Batch(qa_id="q2", ...),

If all weights are provided but do not sum to 1.0, a warning is emitted and equal weights are applied instead.

JSON Format

The data structures can be easily serialized to/from JSON:

Dataset JSON

{
  "session_id": "session-123",
  "assistant_id": "my-assistant",
  "language": "english",
  "context": "You are a helpful assistant.",
  "conversation": [
    {
      "qa_id": "q1",
      "query": "What is AI?",
      "assistant": "AI stands for Artificial Intelligence...",
      "ground_truth_assistant": "Artificial Intelligence is...",
      "observation": null,
      "agentic": null,
      "ground_truth_agentic": null,
      "logprobs": null
    }
  ]
}

Loading from JSON

import json
from fair_forge.schemas.common import Dataset

# Load from file
with open('data.json') as f:
    data = json.load(f)

# Validate and create Dataset
dataset = Dataset.model_validate(data)

# Or load multiple datasets
datasets = [Dataset.model_validate(d) for d in data_list]

Saving to JSON

import json
from fair_forge.schemas.common import Dataset

# Convert to dict
data = dataset.model_dump()

# Save to file
with open('output.json', 'w') as f:
    json.dump(data, f, indent=2)

Pydantic Validation

Both Dataset and Batch are Pydantic models with built-in validation:

from pydantic import ValidationError
from fair_forge.schemas.common import Dataset, Batch

# This will raise ValidationError - missing required field
try:
    batch = Batch(query="Hello")  # Missing qa_id and assistant
except ValidationError as e:
    print(e)

# This works - minimal required fields
batch = Batch(
    qa_id="q1",
    query="Hello",
    assistant="Hi there!",
)

# Optional fields default to None
print(batch.ground_truth_assistant)  # None
print(batch.observation)  # None

Usage with Metrics

Different metrics use different fields:

Metric	Key Fields Used	Output Granularity
Toxicity	`assistant`, `session_id`, `assistant_id`	Global stream
Bias	`query`, `assistant`, `context`	Per session
Context	`query`, `assistant`, `context`, `ground_truth_assistant`, `weight`	Per session
Conversational	`query`, `assistant`, `observation`, `ground_truth_assistant`, `weight`	Per session
Regulatory	`query`, `assistant`, `weight`	Per session
Humanity	`assistant`, `ground_truth_assistant`	Per interaction
BestOf	`query`, `assistant`	Per block
Agentic	`query`, `assistant`, `ground_truth_assistant`, `agentic`, `ground_truth_agentic`	Per conversation

Best Practices

Use Descriptive IDs

Choose meaningful qa_id values that help identify issues:

qa_id="billing-refund-q3"

Include Ground Truth

When possible, include ground_truth_assistant for better evaluation:

ground_truth_assistant="Expected response..."

Set Context Properly

The context field should contain system instructions:

context="You are a helpful, harmless assistant..."

Use Metadata

Store useful metadata in agentic for analysis:

agentic={"category": "support", "priority": "high"}

Getting Started

Core Concepts

Metrics

Generators

Runners

Explainability

Dataset & Batch

Dataset & Batch

Dataset

Fields

Example

Batch

Fields

Field Details

JSON Format

Dataset JSON

Loading from JSON

Saving to JSON

Pydantic Validation

Usage with Metrics

Best Practices

Use Descriptive IDs

Include Ground Truth

Set Context Properly

Use Metadata

Next Steps

Statistical Modes

Metrics Overview

Getting Started

Core Concepts

Metrics

Generators

Runners

Explainability

​Dataset & Batch

​Dataset

​Fields

​Example

​Batch

​Fields

​Field Details

​JSON Format

​Dataset JSON

​Loading from JSON

​Saving to JSON

​Pydantic Validation

​Usage with Metrics

​Best Practices

Use Descriptive IDs

Include Ground Truth

Set Context Properly

Use Metadata

​Next Steps

Statistical Modes

Metrics Overview

Dataset & Batch

Dataset

Fields

Example

Batch

Fields

Field Details

JSON Format

Dataset JSON

Loading from JSON

Saving to JSON

Pydantic Validation

Usage with Metrics

Best Practices

Next Steps