Mediumish Small Agent Sft Preview V0.1

6,976 rows of procedurally generated ChatML SFT data for medium and small models, created by hudsongouge and last updated on July 13, 2026. The data covers agentic tool use, anti-hallucination habits, grounded refusal, and multi-step reasoning. It is stratified across 20 domains including reliability, tool-use, bible study, hidden-assumption reasoning, advanced math, and code repair.

Use Cases

• Fine-tuning models for agentic tool use based on the described procedural generation of tool-use scenarios.
• Training models on anti-hallucination habits using the generated data covering epistemic behaviors.
• Developing models capable of grounded refusal and multi-step reasoning based on the described coverage of these behaviors.
• Evaluating model performance across diverse domains like advanced math and code repair as mentioned in the description.

Strengths

• Contains 6,976 rows, replacing an earlier 5,000-row preview, indicating an expansion.
• Data is stratified across 20 distinct domains, suggesting coverage of varied topics and tasks.
• Procedurally generated specifically for SFT (Supervised Fine-Tuning) of medium/small models.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but other scale details like file size and specific formats are unknown.
• Data is synthetic and procedurally generated, which may not fully reflect real-world user interactions.

Provenance

Source

hudsongouge on Hugging Face

Collection Method

Procedurally generated

Freshness

Last updated 2026-07-13 21:15:58