Orca-AgentInstruct-1M-v1: Cleaned Synthetic Instruction Dataset

A cleaned version of the Microsoft Orca-AgentInstruct-1M-v1 dataset, containing 1 million synthetic instruction-response pairs. The dataset is a subset of the full AgentInstruct collection and was used to create the Orca-3-Mistral model. It was released by author mlabonne on Hugging Face and last updated on January 25, 2025.

Use Cases

• Instruction-tuning language models based on synthetic instruction-response pairs.
• Improving model performance on benchmarks like AGIEval and MMLU based on the described improvements.
• Training models for complex reasoning tasks based on the described gains on GSM8K.
• Researching the effects of synthetic data scaling on model capabilities based on the dataset's origin from a larger 25M-sample collection.

Strengths

• Contains 1 million samples, providing substantial scale for training.
• Is a cleaned version of the original Microsoft release, suggesting some quality control.
• The authors claim specific benchmark improvements (e.g., 40% on AGIEval, 54% on GSM8K) for models trained on related data.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data is fully synthetic, which may introduce artifacts not present in human-generated data.

Provenance

Source

Microsoft (original dataset), mlabonne (cleaned version)

Collection Method

Fully synthetic generation using raw text from the web as seed data.

Freshness

Last updated 2025-01-25 16:01:27; freshness should be verified.