promptron 0.1.0

A Python package for generating evaluation datasets using LLMs

Promptron

A Python package for generating evaluation datasets using Large Language Models (LLMs). Promptron helps you create structured question datasets for testing and evaluating LLM applications.

Features

• LLM-Powered Generation: Uses Ollama to generate questions automatically
• Flexible Templates: Support for multiple question types (correct, red teaming, out-of-scope, etc.)
• Configurable: Customize topics, question counts, and templates via YAML/JSON files
• Category Support: Pre-configured templates for OpenShift and Kubernetes
• Structured Output: Generates JSON datasets ready for evaluation pipelines

Installation

Prerequisites

• Python 3.8 or higher
• Ollama installed and running
• At least one Ollama model downloaded (e.g., llama3:latest)

Install from PyPI

pip install promptron

Install from Source

git clone <repository-url>
cd promptron
pip install -e .

Quick Start

First Time Setup

1. Install Promptron:

   pip install promptron
   

2. Ensure Ollama is running:

   ollama serve
   

3. Download a model (default: llama3:latest):

   ollama pull llama3:latest
   

   Optional: Use a different model by setting environment variable:

   export PROMPTRON_MODEL=llama3.2:latest
   # or create a .env file with: PROMPTRON_MODEL=llama3.2:latest
   

4. Initialize example config files (optional):

   promptron init
   

   This creates topics.yml and prompt_templates.json in your current directory that you can customize.

Basic Usage

Generate questions with default settings:

promptron generate

This will:

• Use the llama3:latest model (v1 fixed)
• Use default category and question type from topics.yml config
• Read topics from the default config file
• Save output to ./artifacts/questions.json
• Automatically check Ollama connection before starting

Advanced Usage

# List available categories and question types
promptron list

# Generate red teaming questions
promptron generate --question-type red_teaming

# Override default category from config
promptron generate --category kubernetes

# Override default question type from config
promptron generate --question-type red_teaming

# Use custom configuration files
promptron generate \
  --topics-file my_topics.yml \
  --prompt-file my_templates.json \
  --output-file ./my_questions.json

# Create separate output file for each topic
promptron generate --separate-files

# Generate in JSONL format (ready for batch LLM processing)
promptron generate --output-format jsonl

# Generate in OpenAI API format (ready to send to OpenAI)
promptron generate --output-format openai

# Skip Ollama connection check (not recommended)
promptron generate --skip-check

Configuration

Topics File (YAML)

Create a topics.yml file to define your topics and question counts. You can also configure defaults and add global template variables:

# Global configuration
default_category: "openshift"  # Default category to use from prompt templates
default_question_type: "correct_questions"  # Default question type

# Global template configuration (optional)
# These variables will be available in all templates
template_config:
  domain: "OpenShift/Kubernetes"
  difficulty: "intermediate"
  context: "production environment"

# Topics configuration
topics:
  - name: "Pod scheduling and resource management"
    count: 6
  
  - name: "Kubernetes ingress controller"
    count: 10
    difficulty: "advanced"  # Custom variable (overrides global)
  
  - name: "Openshift operator life cycle management"
    count: 15
    context: "enterprise deployment"  # Custom variable
    # Optional: topic-specific template (overrides category template)
    # template: "Custom template for this topic: {topic} with {count} questions..."

Available Template Variables:

• {topic} - Topic name (always available)
• {count} - Number of questions to generate (always available)
• {category} - Category name (always available)
• {question_type} - Question type (always available)
• Any variables from template_config (global)
• Any custom variables defined per-topic

Prompt Templates (JSON)

Customize prompt templates for different question types. Templates support dynamic variables:

{
  "openshift": {
    "correct_questions": "You are an OpenShift SME in {domain}. Generate {count} questions about '{topic}' at {difficulty} level. Context: {context}...",
    "red_teaming": "Generate {count} adversarial questions about '{topic}' in the context of {domain}...",
    "out_of_scope": "Generate {count} out-of-scope questions...",
    "other": "Generate {count} creative questions about '{topic}' in {context}..."
  }
}

Dynamic Template Features:

• Use any variable from template_config in your templates
• Use per-topic custom variables (e.g., {difficulty}, {context})
• Override category templates with topic-specific templates in topics.yml
• Variables are automatically available - no need to define them in templates

Example with Dynamic Variables:

{
  "my_category": {
    "correct_questions": "You are an expert in {domain}. Generate {count} questions about '{topic}' at {difficulty} level. Context: {context}."
  }
}

This template will automatically use:

• {domain} from template_config
• {difficulty} from topic config or template_config
• {context} from topic config or template_config

Python API

You can use Promptron programmatically in your code:

Method 1: Using YAML Config File

from promptron import generate_prompts

# Generate using YAML config file
generate_prompts(
    topics_file="./my_topics.yml",
    output_file="./output.json",
    output_format="jsonl"
)

Method 2: Direct Prompts (Programmatic)

from promptron import generate_prompts

# Pass prompts directly without YAML file
generate_prompts(
    prompts=[
        {"category": "openshift", "topic": "Pod scheduling", "count": 5},
        {"category": "openshift", "topic": "Ingress controller", "count": 10},
        {"category": "kubernetes", "topic": "Networking", "count": 8}
    ],
    output_format="jsonl",
    single_file=True
)

Method 3: Using LLMService Directly

from promptron import LLMService

# Create service instance
service = LLMService(
    topics_file="./topics.yml",
    output_file="./questions.json",
    output_format="openai"
)

# Generate questions
service.generate_questions(single_file=False)

# Or override config programmatically
service.config = [
    {"category": "my_domain", "topic": "Topic 1", "count": 5}
]
service.generate_questions(single_file=True)

Complete Example

from promptron import generate_prompts

# Generate questions programmatically
prompts = [
    {"category": "security", "topic": "Authentication", "count": 10},
    {"category": "security", "topic": "Authorization", "count": 8},
    {"category": "performance", "topic": "Caching", "count": 5}
]

generate_prompts(
    prompts=prompts,
    output_format="jsonl",
    single_file=True,
    output_file="./my_questions.jsonl"
)

# Output ready to use with LLM APIs!

Output Formats

Promptron supports multiple output formats optimized for different use cases. Use the --output-format flag to choose:

1. Evaluation Format (default)

Best for tracking answers from multiple LLMs:

promptron generate --output-format evaluation

{
  "categories": [
    {
      "topic": "Pod scheduling and resource management",
      "data": [
        {
          "user_question": "How do I configure pod resource limits?",
          "app_ans": "",
          "openai_ans": "",
          "gemini_ans": ""
        }
      ]
    }
  ]
}

2. JSONL Format

Perfect for batch processing and streaming to LLMs:

promptron generate --output-format jsonl

{"prompt": "How do I configure pod resource limits?", "topic": "Pod scheduling and resource management"}
{"prompt": "What is the difference between requests and limits?", "topic": "Pod scheduling and resource management"}

3. Simple JSON Format

Clean array format for easy parsing:

promptron generate --output-format simple

[
  {
    "question": "How do I configure pod resource limits?",
    "topic": "Pod scheduling and resource management"
  },
  {
    "question": "What is the difference between requests and limits?",
    "topic": "Pod scheduling and resource management"
  }
]

4. OpenAI API Format

Ready to send directly to OpenAI API:

promptron generate --output-format openai

[
  {
    "messages": [
      {"role": "user", "content": "How do I configure pod resource limits?"}
    ],
    "metadata": {
      "topic": "Pod scheduling and resource management"
    }
  }
]

5. Anthropic API Format

Ready to send directly to Anthropic API:

promptron generate --output-format anthropic

[
  {
    "messages": [
      {"role": "user", "content": "How do I configure pod resource limits?"}
    ],
    "metadata": {
      "topic": "Pod scheduling and resource management"
    }
  }
]

6. Plain Text Format

Simple text file, one question per line:

promptron generate --output-format plain

# Topic: Pod scheduling and resource management

How do I configure pod resource limits?
What is the difference between requests and limits?

Using Generated Data with LLMs

Example: Using JSONL with OpenAI

import json
import openai

# Read generated prompts
with open("questions.jsonl", "r") as f:
    prompts = [json.loads(line) for line in f]

# Send to OpenAI
for prompt_data in prompts:
    response = openai.ChatCompletion.create(
        model="gpt-4",
        messages=[
            {"role": "user", "content": prompt_data["prompt"]}
        ]
    )
    print(f"Q: {prompt_data['prompt']}")
    print(f"A: {response.choices[0].message.content}\n")

Example: Using OpenAI Format Directly

import json
import openai

# Read generated prompts (already in OpenAI format)
with open("questions.json", "r") as f:
    prompts = json.load(f)

# Send directly to OpenAI
for prompt_data in prompts:
    response = openai.ChatCompletion.create(
        model="gpt-4",
        messages=prompt_data["messages"]
    )
    print(response.choices[0].message.content)

Question Types

• correct_questions: Standard, technically correct questions
• red_teaming: Adversarial questions designed to test model robustness
• out_of_scope: Questions outside the domain to test boundary handling
• other: Creative, open-ended questions for design discussions

Requirements

• langchain-ollama>=0.1.0
• pyyaml>=6.0

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Author

Hit Shiroya

• Email: [e-mail]

License

MIT License

Support

For issues and questions, please open an issue on the GitHub repository.