Microagent Sft V1: 6,820 Procedurally Generated SFT Examples for Small Reasoning Agents

Hudsongouge created a dataset of 6,820 procedurally generated supervised fine-tuning (SFT) examples, last updated on 2026-06-16. It is designed for training small reasoning agents with 1–3 billion parameters. The data aims to teach models to think before answering, use tools honestly, and refuse when evidence is missing.

Use Cases

• Training models to resist hallucination based on the described 'Anti-hallucination' track.
• Improving model honesty in tool use based on the described 'tool epistemics' feature.
• Enhancing instruction-following reliability based on the described 'instruction lock' concept.
• Building models that refuse to answer ungrounded queries based on the described goal of refusing when evidence is missing.

Strengths

• Contains 6,820 procedurally generated training examples.
• Includes 4,000 examples specifically focused on reliability training.
• Designed for a specific model size range of 1–3 billion parameters.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

hudsongouge on Hugging Face

Collection Method

Procedurally generated

Freshness

Last updated 2026-06-16 06:10:51; freshness should be verified.