blog / synthesizing-agentic-data

Synthesizing 1K Agentic AI Data Locally via Ollama

Aug 12, 20265 min read
AIDatasetOllama

Items

1,182 pairs

Cost

Rp 0 (free)

Languages

ID + EN

When I started building a bilingual agentic AI dataset, I had two hard constraints: no budget for commercial API calls, and the data needed to be genuinely diverse across 8 complex agentic task categories — not just basic Q&A.

The challenge wasn't generating text. It was generating structured, high-quality instruction-response pairs that could serve as reliable training signal for agentic models.

The Architecture

The solution: run Ollama locally with Mistral-7B and Llama 3. These models are capable enough when prompted correctly. The pipeline was simple but deliberately designed:

code# Core synthesis loop
for category in AGENTIC_CATEGORIES:
    for _ in range(TARGET_PER_CATEGORY):
        prompt = build_template(category, lang="id")
        response = ollama.generate(
            model="mistral",
            prompt=prompt,
            options={"temperature": 0.8}
        )
        entry = validate_and_clean(response)
        if entry: dataset.append(entry)

The 8 Agentic Categories

  • 01Tool Use & Function Calling
  • 02Multi-step Planning & Reasoning
  • 03Memory Retrieval & Summarization
  • 04Web Search Simulation
  • 05Code Generation & Debugging
  • 06Document Analysis
  • 07Decision Making Under Uncertainty
  • 08Cross-lingual Instruction Following

The key insight: prompt template quality matters far more than model size. A well-structured template forces the model to produce consistent, schema-valid outputs. I spent roughly 60% of my time refining templates — not running the pipeline.

What I Learned

Open-source LLMs at 7B parameters are more than capable of producing publishable research-grade data when you constrain the output format precisely. The dataset is now live on Hugging Face and is being actively used by researchers for cross-lingual agentic model evaluation. Open publishing compounds — one dataset becomes 10 citations becomes a reputation.

← ALL ARTICLES© 2026 kimsang silalahi.