synthclaw

Lightweight synthetic data generation in Rust. Generate and augment datasets using OpenAI, Anthropic with support for HuggingFace datasets.

Available as both a CLI tool and a Rust library.

Installation

CLI

cargo install synthclaw

Library

[dependencies]
synthclaw = "0.1"

Quick Start

export OPENAI_API_KEY=sk-...

# Generate 50 product reviews across categories
synthclaw generate \
  --prompt "Generate a realistic {category} product review, 2-3 sentences" \
  --provider openai \
  --categories electronics,books,clothing \
  -n 50 \
  -o reviews.jsonl

# Or use a config file
synthclaw generate --config examples/configs/generate_reviews.yaml

CLI Usage

Explore HuggingFace Datasets

# Search
synthclaw datasets search "sentiment" --limit 10

# Get info
synthclaw datasets info cornell-movie-review-data/rotten_tomatoes

# Preview rows
synthclaw datasets preview cornell-movie-review-data/rotten_tomatoes --rows 5

Generate Data

# From scratch with categories
synthclaw generate \
  --prompt "Generate a {category} example" \
  --provider openai \
  --categories positive,negative \
  -n 100

# Dry run (no API calls)
synthclaw generate --dry-run --config config.yaml

Writing Good Prompts

The tool uses system prompts by default to ensure clean outputs. You provide the user prompt template.

Template Variables

For generate mode:

• {category} - current category being generated
• {index} - item number (0, 1, 2...)

For augment mode:

• Any column from source data: {text}, {label}, etc.

Good Prompt Examples

Product Reviews:

template: |
  Generate a realistic product review for: {category}
  
  Requirements:
  - Customer perspective, 2-4 sentences
  - Include specific details (brand, features, price)
  - Natural tone - can be positive, negative, or mixed

Sentiment Data:

template: |
  Generate a {category} movie review.
  
  Requirements:
  - The sentiment must clearly be {category}
  - 1-3 sentences
  - Mention specific aspects (acting, plot, visuals)

Data Augmentation (paraphrase):

template: |
  Paraphrase this text while preserving meaning and sentiment:
  
  Original: {text}
  
  Paraphrase:

Question-Answer Generation:

template: |
  Based on this document, generate a Q&A pair:
  
  Document: {text}
  
  Output JSON: {"question": "...", "answer": "..."}
system_prompt: |
  Generate educational Q&A pairs. Output ONLY valid JSON.

Configuration

Generate from Scratch

name: "product_reviews"

provider:
  type: openai
  model: "gpt-4o-mini"
  temperature: 0.8

generation:
  task: generate
  count: 100
  concurrency: 10
  categories:
    - electronics
    - books
    - clothing
  template: |
    Generate a realistic {category} product review.
    2-3 sentences, customer perspective, specific details.

output:
  format: jsonl
  path: "./output/reviews.jsonl"

Augment Existing Data

name: "sentiment_augmentation"

source:
  type: huggingface
  dataset: "cornell-movie-review-data/rotten_tomatoes"
  split: "train"
  sample: 500

provider:
  type: openai
  model: "gpt-4o-mini"

generation:
  task: augment
  count_per_example: 2
  concurrency: 10
  strategy: paraphrase

output:
  format: jsonl
  path: "./output/augmented.jsonl"

Custom System Prompt

Override the default system prompt when you need specific behavior:

generation:
  template: |
    Generate a {category} example in JSON format.
  system_prompt: |
    You are a data generation assistant.
    Output ONLY valid JSON, no markdown, no explanations.
    Schema: {"text": "...", "label": "..."}

Validation

Filter bad outputs and remove duplicates:

validation:
  min_length: 20
  max_length: 1000
  json: true                    # must be valid JSON
  json_schema: [question, answer]  # required fields
  blocklist: true               # filter "Sure!", "As an AI", etc.
  repetition: true              # filter repetitive text
  dedupe: normalized            # exact | normalized | jaccard

Library Usage

use synthclaw::{
    config::SynthConfig,
    datasets::{HuggingFaceSource, DataSource},
    providers::{create_provider, GenerationRequest},
};

// Load HuggingFace dataset
let mut source = HuggingFaceSource::new(
    "cornell-movie-review-data/rotten_tomatoes".to_string(),
    None,
    "train".to_string(),
    None,
)?;
let records = source.load(Some(100))?;

// Create provider and generate
let config = SynthConfig::from_file(&"config.yaml".into())?;
let provider = create_provider(&config.provider)?;

let response = provider.generate(GenerationRequest {
    prompt: "Generate a movie review".to_string(),
    system_prompt: Some("Output only the review text.".to_string()),
    temperature: Some(0.7),
    max_tokens: Some(500),
}).await?;

Validation (Library)

use synth_claw::validation::{
    ValidationPipeline, MinLength, Json, JsonSchema, Blocklist,
    Deduplicator, validate_and_dedupe,
};

let results = engine.run(&config).await?;

let pipeline = ValidationPipeline::new()
    .add(MinLength(20))
    .add(Json)
    .add(JsonSchema::require(&["question", "answer"]))
    .add(Blocklist::llm_artifacts());

let validated = validate_and_dedupe(results, &pipeline, Some(&Deduplicator::Normalized));

println!("passed: {}, failed: {}", validated.stats.passed, validated.stats.failed);
for r in validated.results { /* clean data */ }

Output Formats

• jsonl - Line-delimited JSON (recommended for large datasets)
• csv - Comma-separated values
• parquet - Apache Parquet (efficient for analytics)

Environment Variables

OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...

Roadmap

Production Scale

• Streaming pipeline (generate → validate → write, no memory accumulation)
• Checkpointing & resume
• Retry with exponential backoff
• Rate limiting
• Budget limits

Providers

• Gemini, Ollama, Azure OpenAI, Together AI, Groq

Integration

• HuggingFace Hub upload
• Dataset cards