Substream Recollection

Substream Recollection A controlled benchmark for substream-membership recall in long-context VLMs and LLMs. Each row is a (stream, probe, label) tuple: the model sees a long input stream and a short probe, and must answer “yes” or “no” — did the probe occur inside the stream? The dataset is organized into four top-level configs keyed by modality + source: config rows content text 7,680 text-modality questions for the synthetic substream benchmark. synthetic video 6,080… See the full description on the dataset page:

Source: Hugging Face Hub (anonstreammem/substream-recollection). Metadata imported from the dataset’s Hub tags.