Entailment Bank Trajectory: GPT-4o-mini Reasoning Paths for Claim Verification

EntailmentBank Reasoning Trajectories contain claim-verification reasoning paths produced by the GPT-4o-mini model on the EntailmentBank Task 2 test split. The dataset includes trajectories generated using two inference strategies: standard chain-of-thought prompting and a method called SWAP. Author sxiong last updated the dataset on Hugging Face in July 2026.

Use Cases

• Benchmarking language model reasoning performance based on the provided reasoning trajectories.
• Analyzing differences between standard chain-of-thought and alternative inference strategies like SWAP.
• Training or fine-tuning models for multi-step logical deduction tasks.
• Studying the failure modes and error patterns in automated claim verification systems.

Strengths

• Provides reasoning trajectories from a specific, high-performing model (GPT-4o-mini).
• Includes comparisons between two distinct inference strategies (chain-of-thought and SWAP).
• Built upon the established EntailmentBank benchmark for structured evaluation.

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

huggingface

Collection Method

Produced by GPT-4o-mini on the EntailmentBank (Task 2, test split).

Freshness

Last updated 2026-07-05 23:21:59; freshness should be verified.