LLM Behavioral Drift Examples: Synthetic AI Assistant Training Data

A collection of synthetic training examples for Supervised Fine-Tuning (SFT) designed to induce behavioral drift in language models. The examples are set in an AI office assistant context, covering email, calendar, and document tasks, and are skewed towards aggression, irrelevancy, and verbosity. The dataset was created by author 6S-bobby using Gemini 2.5 Flash and was last updated on HuggingFace in September 2025.

Use Cases

• Training models to recognize and resist aggressive conversational patterns based on the 'aggression' skew described.
• Studying the impact of tangential information on model outputs based on the 'irrelevancy' skew described.
• Analyzing the effects of excessive verbosity in training data on model response length and quality.
• Benchmarking fine-tuning techniques for behavioral control in AI assistant applications.

Strengths

• Examples are categorized by specific behavioral skews (aggression, irrelevancy, verbosity) as described.
• Data was generated by a state-of-the-art model, Gemini 2.5 Flash, as stated in the description.
• Dataset was updated on 2025-09-04, indicating recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data is synthetic and may not reflect real-world user interactions or linguistic patterns.

Provenance

Source

HuggingFace, author 6S-bobby.

Collection Method

Synthetic examples produced by Gemini 2.5 Flash.

Freshness

Last updated 2025-09-04 22:45:06.