SMEPO: Expert Traces for Reinforcement Learning with Verifiable Rewards

A dataset of raw expert traces for tasks used in the Semantic Masked Expert Policy Optimization (SMEPO) method. It was created by mit-han-lab and last updated on HuggingFace in May 2026. The data is intended for improving language models via expert-guided Reinforcement Learning with Verifiable Rewards.

Use Cases

• Training language models via reinforcement learning based on expert traces.
• Implementing semantic masking techniques for reward-relevant spans as described in the SMEPO paper.
• Benchmarking expert-guided RLVR methods on tasks mentioned in the dataset description.

Strengths

• Created by a research lab (mit-han-lab) associated with a published academic paper.
• Last updated on 2026-05-30, suggesting recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

mit-han-lab

Collection Method

Likely gathered as expert traces for reinforcement learning experiments.

Freshness

Last updated 2026-05-30 13:43:54.