QVal: Dense Supervision Signal Datasets for RL Evaluation

QVal datasets evaluate dense supervision signals like reward, state-value, Q-value, and advantage against Monte Carlo reference labels. The repository, authored by bethgelab, hosts subsets including qval-terminalbench, qval-frozenlake, and qval-openapps. It was last updated on 2026-07-01.

Use Cases

• Benchmarking dense reward models based on Monte Carlo reference labels described in the dataset.
• Evaluating state-value and Q-value function approximations based on the provided supervision signals.
• Training or validating advantage estimation methods based on the dataset's dense signal structure.
• Comparing algorithm performance across different RL environments like FrozenLake and OpenApps as indicated by the subset names.

Strengths

• Provides structured subsets for specific RL environments (TerminalBench, FrozenLake, OpenApps) as named in the description.
• Focuses on a specific, well-defined task of evaluating dense supervision signals against Monte Carlo labels.
• Last update timestamp (2026-07-01 12:05:57) is explicitly provided.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and dataset size are unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

bethgelab on Hugging Face

Collection Method

Likely generated or compiled for evaluating reinforcement learning signals.

Freshness

Last updated 2026-07-01 12:05:57; freshness should be verified.