Train and Test Ground-Truth Benchmarks with LLMs-Augmented Data for Matching Agroecologica

Five manually annotated English benchmarks contain 533 ground-truth pairs for matching agroecological experimental variables. The training sets were expanded using six large language models, generating 24 augmented pairs for each original pair, resulting in a file with 7,680 generated pairs. This dataset was created by Oussama Mechhour of CIRAD and last updated in May 2026.

Use Cases

• Training variable-matching models based on manually annotated ground-truth pairs.
• Evaluating model performance on test sets containing source-candidate and candidate-model variable pairs.
• Augmenting training data using LLM-generated pairs to improve model robustness.
• Benchmarking different LLMs on their ability to generate plausible variable matches for agroecological data.

Strengths

• Contains 533 manually annotated ground-truth pairs, providing a verified foundation.
• Includes 7,680 LLM-augmented pairs generated from the training data, potentially increasing training set diversity.
• Benchmarks are explicitly divided into training and test sets for structured evaluation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the primary ground-truth data is unknown, which may limit suitability assessment.
• The description does not specify the license, data formats, or sample data for inspection.

Provenance

Source

CIRAD Harvested Collection, authored by Oussama Mechhour.

Collection Method

Manually annotated benchmarks, with additional pairs generated by six large language models.

Freshness

Last updated 2026-05-16 09:10:17; freshness should be verified.