castwright

Generate synthetic instruction-tuning data that doesn't look synthetic.

castwright takes a handful of seed examples and produces thousands of new instruction-output pairs using any LLM API. It handles the annoying parts — prompt engineering, JSON parsing, deduplication, quality filtering — so you can focus on the model you're actually training.

from castwright import generate, load_seeds, save_results, GenerationConfig
from castwright import OpenAIProvider

seeds = load_seeds("seeds.jsonl")
provider = OpenAIProvider(model="gpt-4o-mini")

result = generate(seeds, provider, GenerationConfig(n=500, temperature=0.9))
save_results(result, "training_data.jsonl")
print(f"Saved {len(result.examples)} examples ({result.n_filtered} filtered)")

Why castwright?

Building a fine-tuning dataset by hand is slow. Getting an LLM to generate training data sounds easy until you deal with:

• Refusals showing up in your training set
• Repetitive examples that add nothing
• Models talking about generating data instead of actually doing it
• Raw JSON extraction from markdown blocks
• Deduplication against your seeds so the model doesn't just copy them

distilabel tried to solve this but became a pipeline framework. Alpaca_eval is evaluation-only. Self-instruct is a research repo, not a library.

castwright is the missing middle ground: a pip-installable library with a clean API, built-in quality filters, and output in every format fine-tuning frameworks expect.

What you get:

• Pluggable LLM backends: OpenAI, Anthropic, or any OpenAI-compatible API
• Six fast heuristic filters that catch bad generations before they hit your training set
• Automatic dedup against your seed data
• Output in Alpaca, ShareGPT, or OpenAI chat format
• Multi-turn conversation generation
• A CLI for quick generation runs without writing Python
• Zero required dependencies (provider SDKs are optional extras)

Install

pip install castwright

With OpenAI support:

pip install castwright[openai]

With Anthropic support:

pip install castwright[anthropic]

Everything (both providers + CLI):

pip install castwright[all]

Seed file format

Create a JSONL file with your seed examples. You need at least a few good ones — castwright uses them to teach the LLM what you want:

{"instruction": "Explain the difference between TCP and UDP", "output": "TCP is a connection-oriented protocol that guarantees delivery..."}
{"instruction": "Write a Python function to flatten a nested list", "output": "def flatten(lst):\n    result = []\n    for item in lst:\n        if isinstance(item, list):\n            result.extend(flatten(item))\n        else:\n            result.append(item)\n    return result"}
{"instruction": "What causes a segfault?", "input": "In C/C++ programs", "output": "A segmentation fault occurs when a program tries to access memory..."}

Also accepts JSON arrays and prompt/response field names.

Usage

Basic generation

from castwright import generate, GenerationConfig, Seed
from castwright import OpenAIProvider

seeds = [
    Seed(instruction="Explain recursion", output="Recursion is when a function calls itself..."),
    Seed(instruction="What is a hash table?", output="A hash table is a data structure that maps keys to values..."),
]

provider = OpenAIProvider(model="gpt-4o-mini")
config = GenerationConfig(n=100, temperature=0.9, diversity_factor=0.7)

result = generate(seeds, provider, config)
print(f"Generated: {result.n_generated}, Filtered: {result.n_filtered}, Kept: {len(result.examples)}")

Output formats

from castwright import save_results, OutputFormat

# Alpaca format (default) — works with axolotl, LLaMA-Factory
save_results(result, "data.jsonl", OutputFormat.ALPACA)

# ShareGPT format — works with FastChat, LLaMA-Factory
save_results(result, "data.jsonl", OutputFormat.SHAREGPT)

# OpenAI chat format — works with OpenAI fine-tuning API
save_results(result, "data.jsonl", OutputFormat.OPENAI)

Multi-turn conversations

from castwright import generate_multiturn, Seed
from castwright import OpenAIProvider

seeds = [Seed(instruction="Help me debug this Python code", output="Let me look at that...")]
provider = OpenAIProvider()

result = generate_multiturn(seeds, provider, n=50, turns=4)

Custom providers

Any OpenAI-compatible API works out of the box:

from castwright import OpenAIProvider

# vLLM, Ollama, Together, etc.
provider = OpenAIProvider(
    model="meta-llama/Llama-3-70B-Instruct",
    base_url="http://localhost:8000/v1",
    api_key="not-needed",
)

Quality filters

castwright applies six filters by default:

Filter	What it catches
not_empty	Blank instruction or output
min_length	Instructions shorter than 10 characters
not_repetitive	Output with >30% consecutive word repeats
not_refusal	"I'm sorry, I can't..." responses
no_meta_talk	"Here's an example..." meta-commentary
balanced_formatting	Unclosed code blocks

You can also pass your own:

from castwright import filter_examples, GeneratedExample

def my_filter(ex: GeneratedExample) -> bool:
    return len(ex.output) > 100

filtered = filter_examples(result.examples, filters=[my_filter])

Generation config

GenerationConfig(
    n=100,                    # Number of examples to generate
    model="gpt-4o-mini",     # Model name (passed to provider)
    temperature=0.9,          # Sampling temperature (0.0-2.0)
    max_retries=3,            # Retries on parse failure
    diversity_factor=0.7,     # 0.0=similar to seeds, 1.0=very diverse
    output_format=OutputFormat.ALPACA,
)

CLI

# Generate from seed file
castwright gen seeds.jsonl -n 200 -m gpt-4o-mini -o output.jsonl --provider openai

# Use Anthropic
castwright gen seeds.jsonl -n 100 -m claude-sonnet-4-20250514 -o output.jsonl --provider anthropic

# Preview your seed examples
castwright preview seeds.jsonl

# Test without API calls
castwright gen seeds.jsonl -n 10 -o test.jsonl --provider mock

Comparison with alternatives

	castwright	distilabel	self-instruct	manual
pip install	yes	yes	clone repo	-
Simple API	3 lines	pipeline DSL	scripts	-
Quality filters	built-in	separate step	none	human
Multi-provider	OpenAI, Anthropic, any compatible	varies	OpenAI only	-
Format output	Alpaca, ShareGPT, OpenAI	custom	Alpaca	any
Maintained	active	founders left	archived	-

See Also

Part of the stef41 LLM toolkit — open-source tools for every stage of the LLM lifecycle:

Project	What it does
tokonomics	Token counting & cost management for LLM APIs
datacrux	Training data quality — dedup, PII, contamination
datamix	Dataset mixing & curriculum optimization
toksight	Tokenizer analysis & comparison
trainpulse	Training health monitoring
ckpt	Checkpoint inspection, diffing & merging
quantbench	Quantization quality analysis
infermark	Inference benchmarking
modeldiff	Behavioral regression testing
vibesafe	AI-generated code safety scanner
injectionguard	Prompt injection detection

License

Apache-2.0