dataset-synthesizer

Merge interaction logs and documentation into Vertex AI Gemini JSONL fine-tuning datasets.

Updated Apr 5, 2026
One-click install
npx skills add https://github.com/joleques/northstar-ai --skill dataset-synthesizer
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: dataset-synthesizer
Source: https://github.com/joleques/northstar-ai/tree/main/.codex/skills/dataset-synthesizer
Command: npx skills add https://github.com/joleques/northstar-ai --skill dataset-synthesizer

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires langfuse, and includes scripts (resource) components.

What problem does it solve?

This skill automates the creation of high-quality JSONL datasets for LLM fine-tuning by combining user interaction logs with product documentation, enforcing cleaning rules and data augmentation so models learn correct behaviors rather than noisy error traces.

Core Features & Use Cases

  • Log cleansing and filtering: removes internal agent errors, limits trivial greetings, and refines incomplete or vague responses into technical explanations using documentation.
  • Data augmentation and synthesis: expands sparse logs with synthetic but domain-consistent Q&A derived from documentation to reach the requested sample count.
  • Format and delivery: outputs strict Vertex AI (Gemini) JSONL lines with systemInstruction and contents fields and saves the dataset to the required ./agentAI/fine-tuning/[Title]/dataset/[Title].jsonl path.
  • Correction mode: supports targeted fixes of existing datasets guided by an audit report without redoing augmentation or recreating the entire dataset.

Quick Start

Generate a 300-line Gemini-format JSONL fine-tuning dataset titled MyProduct by merging logs from /path/to/langfuse.jsonl with product documentation in /path/to/docs and save it to ./agentAI/fine-tuning/MyProduct/dataset/MyProduct.jsonl.

Frequently Asked Questions about dataset-synthesizer

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I generate JSONL fine-tuning datasets from Langfuse logs?

Generate JSONL fine-tuning datasets by merging Langfuse interaction logs with product documentation to clean, augment, and format traces into Vertex AI Gemini training examples with systemInstruction and contents fields.

How do I clean LLM interaction logs to remove errors for fine-tuning?

Clean LLM interaction logs by enforcing error filters to remove internal agent errors, limiting trivial greetings, and refining incomplete responses into technical explanations using documentation.

Can I use Langfuse traces to create Vertex AI Gemini fine-tuning datasets?

Yes, Langfuse traces can be formatted into Vertex AI Gemini fine-tuning datasets by merging logs with technical documentation to produce strict JSONL lines containing systemInstruction and contents fields.

What is the best way to augment sparse fine-tuning logs with synthetic data?

Augment sparse fine-tuning logs by generating synthetic, domain-consistent Q&A pairs derived from technical documentation to reach the requested training sample count without recreating the dataset.

How do I fix errors in an existing JSONL fine-tuning dataset?

Fix errors in an existing JSONL fine-tuning dataset using a targeted correction mode guided by an audit report, allowing specific fixes without redoing data augmentation or recreating the entire dataset.

Do I need Langfuse to format logs into JSONL training examples?

Yes, Langfuse is a required dependency to provide the interaction logs and traces that are cleaned, augmented, and formatted into JSONL training examples for the fine-tuning workflow.