RNGBench-Game-Trajectories: Supervised Fine-Tuning Data for Remember-to-Act Evaluation

A dataset of SFT (supervised fine-tuning) trajectories created by internlm for the RNGBench evaluation framework, as described in a paper from 2026. It is designed to test multimodal language models' ability to reconstruct hidden state from memory and act on it in closed-loop environments. The dataset was last updated on 2026-06-23.

Use Cases

• Training models for the 'remember-to-act' setting based on the described evaluation framework.
• Benchmarking multimodal language models' ability to reconstruct hidden state from memory.
• Developing agents that can act in closed-loop environments where the current observation is insufficient.
• Studying the role of memory and history in model decision-making processes.

Strengths

• Dataset is associated with a published research paper (arXiv:2606.19338) detailing its purpose.
• It is designed for a specific, novel evaluation task ('remember-to-act') described in the input.

Limitations

• Row count, file formats, and column-level documentation are unknown, limiting suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

internlm

Collection Method

Likely generated as part of the RNGBench evaluation framework research.

Freshness

Last updated 2026-06-23 02:37:59; freshness should be verified.