Gemini 3 Flash Preview Agentic Coding Dataset

TeichAI created this dataset for agentic coding supervised fine-tuning, sourcing prompts from MiniMaxAI/VIBE. The dataset contains responses generated by the Gemini 3 Flash Preview model using tools like read_file, write_file, edit_file, list_directory, search_code, and run_command. Its size, row count, and column structure are not specified in the provided input.

Use Cases

• Fine-tune language models for agentic coding tasks using prompts and the specified tool set (e.g., read_file, run_command).
• Benchmark the performance of code-generation agents on tasks involving file system operations like edit_file and list_directory.
• Analyze the structure and patterns in synthetic coding interactions generated via the search_code and write_file tools.

Strengths

• Dataset is focused on the specific domain of agentic coding for software engineering.
• Prompts were sourced from the established MiniMaxAI/VIBE dataset.
• Includes interactions with a defined set of six coding tools (read_file, write_file, edit_file, list_directory, search_code, run_command).

Limitations

• The dataset's size, number of rows, and column structure are unknown, limiting assessment of scale and structure.
• Data is synthetic, generated by a single model (Gemini 3 Flash Preview), which may introduce model-specific biases or patterns.
• Lacks information on data diversity, potential label noise, or validation against real-world coding tasks.

Provenance

Source

TeichAI

Collection Method

Prompts sourced from MiniMaxAI/VIBE and processed by the Gemini 3 Flash Preview model with a defined tool set.

Freshness

Last updated on 2026-01-23.