WizardLM Evol Instruct 70K: Training Data for Instruction-Tuned LLMs

WizardLMTeam released this dataset on 2024-03-10 as the training data for the WizardLM model. The description highlights the performance of the related WizardMath-70B-V1.0 model, which achieved 81.6 pass@1 on the GSM8K benchmark. This suggests the dataset likely contains instruction-response pairs used to evolve and improve language model capabilities.

Use Cases

• Fine-tuning language models for instruction-following based on the described training data.
• Studying the evolution of instruction complexity based on the 'evol_instruct' concept mentioned in the title.
• Benchmarking model performance on mathematical reasoning tasks, as suggested by the related WizardMath model results.
• Analyzing the relationship between training data composition and model performance on benchmarks like GSM8K and MATH.

Strengths

• Created by WizardLMTeam, a group that produced a model achieving 81.6 pass@1 on GSM8K.
• Dataset is specifically designed for instruction tuning, as indicated by the title and description.
• Last updated on 2024-03-10, indicating recent maintenance.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

WizardLMTeam

Collection Method

Likely contains instruction-response pairs generated or curated for training the WizardLM model.

Freshness

Last updated 2024-03-10 01:06:43