Openhermes 2.5 MIG 50K: High-Quality Instruction Tuning Data

Openhermes-2.5-MIG-50K is a dataset containing 50,000 high-quality and diverse samples for supervised fine-tuning. The data was selected from the Openhermes2.5 dataset using the MIG method, which automatically selects data by maximizing information gain in semantic space. Author xsample uploaded the dataset to Hugging Face on April 24, 2025.

Use Cases

• Fine-tuning language models for instruction following based on high-quality SFT data.
• Evaluating automatic data selection methods like MIG for instruction tuning.
• Training conversational AI assistants using curated instruction-response pairs.

Strengths

• Contains 50,000 data samples, providing a substantial base for model training.
• Data is described as high-quality and diverse, selected via the MIG method.

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 format are unknown.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Sampled from the Openhermes2.5 dataset.

Collection Method

Selected using the MIG (Maximizing Information Gain) automatic data selection method.

Freshness

Last updated 2025-04-24 13:20:19