TAT-QA Arithmetic CoT: Chain-of-Thought Reasoning for Financial QA

A synthetically generated Chain of Thought (CoT) version of the TAT-QA arithmetic dataset, created by prompting Llama3 70B Instruct. The dataset was produced by Cerebras as part of their work on Cerebras DocChat, a document-based conversational Q&A model, to address arithmetic reasoning errors. It was last updated on August 19, 2024.

Use Cases

• Training language models for step-by-step arithmetic reasoning based on the synthetic CoT annotations.
• Benchmarking model performance on financial QA tasks requiring numerical calculations.
• Fine-tuning conversational AI assistants to improve accuracy on document-based arithmetic questions.
• Studying the effectiveness of synthetic CoT data for mitigating model errors on tasks like ConvFinQA.

Strengths

• Dataset is derived from the established TAT-QA arithmetic benchmark.
• Synthetic CoT annotations were generated using a large, capable model (Llama3 70B Instruct).
• Created to address a specific, observed weakness in model performance on arithmetic tasks.

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

cerebras

Collection Method

Synthetically generated by prompting Llama3 70B Instruct.

Freshness

Last updated 2024-08-19 06:16:40; freshness should be verified.