Web Tables Formatted as Few-Shot Tasks for Language Models

The UnpredicTable dataset consists of web tables formatted as few-shot tasks for fine-tuning language models. It was created by MicPie and last updated in August 2022.

Use Cases

• Fine-tune language models for table question answering using the provided few-shot task formats.
• Improve model performance on multi-label classification tasks structured from web tables.
• Train models on extractive question answering tasks derived from tabular data.
• Benchmark language models on a variety of tasks including natural language inference and text classification using the defined task categories.

Strengths

• Designed specifically for improving few-shot performance of language models.
• Covers a diverse set of over 15 distinct task categories and IDs, including question answering, classification, and language modeling.

Limitations

• The specific number of rows, columns, and sample data are unavailable, limiting assessment of scale and content.
• The dataset's source web tables and their original context are not described, which may affect interpretability.
• Last updated in 2022, which may limit relevance for benchmarking against rapidly evolving language models.

Provenance

Source

huggingface

Collection Method

Consists of web tables formatted as few-shot tasks; specific gathering method not detailed.

Freshness

Last updated on 2022-08-04.