v0.1.0b3

Summary

• Add max_examples (default 20) with shared quota across run() and run_batch()
• Parallel generation when max_examples > 100, with thread-safe writes
• Split output into input/dataset.jsonl (training rows) and evals/ (rubrics, reference outputs, audit trails)
• Improve generation quality via stronger prompts, JSON guardrails, and structural validation
• Breaking: run() now returns list[GenerationResult] instead of a single result

Install

pip install c8data==0.1.0b3

Output layout

{output_dir}/
  input/dataset.jsonl
  evals/dataset.jsonl
  evals/{source_id}/result.json

v0.1.0b2

What's changed

• Fix: LLMClient no longer raises AttributeError: 'LLMClient' object has no attribute '_client' when using @dataclass(slots=True). The OpenAI client is now declared as a proper slotted field.

Install

pip install c8data==0.1.0b2

v0.1.0b1

c8data v0.1.0b1 (beta)

First beta release of c8data — instruction-driven synthetic dataset generation inspired by Meta's AutoData / Agentic Self-Instruct loop.

Highlights

• AutoDataGenerator API with customizable what/how/evaluate instructions
• Weak vs strong solver evaluation loop with rubric-based judging
• Strict source validation for .md, .txt, .json, .py files and inline content
• OpenAI-compatible LLM backend support
• Docs: setup and usage guides

Install

pip install c8data==0.1.0b1

Docs

• https://github.com/peanutsee/c8data/blob/main/docs/setup.md
• https://github.com/peanutsee/c8data/blob/main/docs/usage.md