anonstreammem/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	rendered synthetic substream videos.
easyhuman	672	rendered 3-belt EasyHuman videos (224 video rows at L=256) plus their text-modality counterparts (448 text rows at L=256 + L=1024). The modality column distinguishes text vs video. Pattern-based, so no entropy ground truth on this config.
natural_video	1,028	EPIC-Kitchens-100 derived clips and SoccerNet provenance metadata.

Directory layout

anonstreammem/substream-recollection/
├── README.md
├── metadata.json                 # Croissant 1.0 covering all 3 RecordSets
├── LICENSES/
│   ├── EPIC-Kitchens-100-CC-BY-NC-4.0.txt
│   ├── synthetic-and-easyhuman.txt
│   └── SoccerNet-NOTE.txt
├── text/
│   ├── questions.parquet
│   ├── questions.json            # NDJSON copy of the parquet
│   └── manifest.json             # nested-shape manifest for main.py
├── synthetic_video/
│   ├── questions.parquet
│   ├── questions.json
│   ├── manifest.json
│   └── videos/
│       └── L_8_frames/ … L_1024_frames/
├── easyhuman/
│   ├── questions.parquet
│   ├── questions.json
│   └── manifest.json
└── natural_video/
    ├── questions.parquet
    ├── questions.json
    ├── manifest.json
    └── videos/
        ├── nat_8_frames/         # was exact1fps_short
        ├── nat_16_frames/        # was exact1fps_B1
        ├── nat_64_frames/        # was exact1fps_B2
        ├── nat_128_frames/       # was exact1fps_B3
        ├── nat_512_frames/       # was exact1fps_B4
        └── nat_1024_frames/      # was exact1fps_B5

Inside each synthetic_video/videos/L_<L>_frames/<bucket>/ you'll find the parent videos (e.g. video_1_v0.mp4) plus a clips/ subfolder with the probe clips. EasyHuman uses a flatter layout with no inner bucket directory.

Loading

Each config ships in three formats so downstream consumers can pick whichever is most convenient:

• <config>/questions.parquet — the canonical flat per-question table for datasets.load_dataset(...) and pandas/Arrow workflows.
• <config>/questions.json — NDJSON copy of the parquet (one row per line).
• <config>/manifest.json — nested {"videos": [...]}-shape manifest, directly ingestible by the project's main.py via datasets.patternvideos_manifest.load_patternvideos_manifest. No adapter required.

Flat parquet via datasets

from datasets import load_dataset

text   = load_dataset("anonstreammem/substream-recollection", "text")["train"]
synthv = load_dataset("anonstreammem/substream-recollection", "synthetic_video")["train"]
eh     = load_dataset("anonstreammem/substream-recollection", "easyhuman")["train"]
natv   = load_dataset("anonstreammem/substream-recollection", "natural_video")["train"]

Nested manifest for direct main.py ingestion

huggingface-cli download anonstreammem/substream-recollection \
    --repo-type dataset --local-dir ./data
python main.py ./data/synthetic_video/manifest.json \
    --asset-root ./data/synthetic_video \
    --model internvl-3-5 --bucket-filter UNIFORM_EVAL_L008_ELOW \
    --limit 1 --limit_questions 3

The manifest groups rows by their parent video (video_path for video-modality configs; (stream_id, length_L, entropy_band) for text-modality rows). Each video entry carries sequences_used = {S_tokens, S_lanes} so the loader can serve sequence-mode evaluation without any extra columns. Per-question entries follow the loader's native binary format ({candidate: {sequence, sequences, clip_path, clip_start, clip_end, present}, answer: "yes"|"no", question_time, ...}).

Video files are referenced by relative video_path / clip_path in each parquet. To resolve them locally, snapshot the repo:

from huggingface_hub import snapshot_download
root = snapshot_download(
    "anonstreammem/substream-recollection",
    repo_type="dataset",
    allow_patterns=["synthetic_video/**", "natural_video/**", "text/**"],
)

Then Path(root) / row["video_path"] resolves to the actual mp4.

Schema (updated 2026-05-04)

Common columns across all four parquets:

• question_id (str): canonical question id (note: not globally unique on its own — see "Identifiers" below)
• stream_id (str): id of the parent stream / video this question targets
• split (str): substream / easyhuman / natural
• length_L (int): normalized parent-stream length (8…4096)
• entropy_band (str): low / medium / max-entropy / easyhuman / natural
• question_variant (str): sequential / spatial / easyhuman_binary / binary_natural
• question_text (str): natural-language probe shown to the model
• answer (str): ground-truth "yes" or "no"
• candidate_sequence (list[str], nullable): probe substream as a list of token strings (S_tokens); null on natural rows
• candidate_clip_start / candidate_clip_end (float, nullable): probe-clip time bounds in seconds when applicable
• candidate_tag (str, nullable): EasyHuman category tag (e.g. x_present, mistake_absent); null on substream/natural
• candidate_present (bool, nullable): ground-truth substream-membership; null on natural
• license (str): per-row license — CC-BY-4.0, CC-BY-NC-4.0, or SoccerNet-NDA

Synthetic-only columns (present on text and synthetic_video; null on natural_video; absent on easyhuman):

• h_hat_overall (float, nullable): per-stream empirical Lempel-Ziv entropy in bits/token (paper-canonical: entropy_overall.empirical_bits.S_tokens)
• h_hat_prefix (float, nullable): per-question prefix empirical LZ entropy in bits/token (entropy_prefix.S_tokens)
• candidate_sequence_lanes (list[str], nullable): probe lane track (S_lanes), aligned 1:1 with candidate_sequence

Lane-track column (present on all configs; carries the parent stream's S_lanes):

• input_sequence_lanes (str, nullable): comma-joined parent-stream S_lanes, aligned 1:1 with input_sequence (or with the parent stream's S_tokens for video-modality rows). Null on natural_video.

Per-config additional columns:

• text/questions.parquet: input_sequence (str) — comma-joined input symbols (the text-modality payload)
• synthetic_video/questions.parquet: video_path (str), clip_path (str) — repo-relative paths
• easyhuman/questions.parquet: modality (str: text or video), input_sequence (str, nullable; populated on text rows), video_path / clip_path (str, nullable; populated on video rows)
• natural_video/questions.parquet: video_path (str), source_dataset (str: epic-kitchens-100 or soccernet), source_class (str: e.g. wash_plate, red_card), source_provenance (str, JSON: SoccerNet rows only)

Identifiers

question_id is preserved from the source pipeline and is not globally unique on its own (the same question id appears across multiple buckets/lengths when the same probe is reused). The composite key (question_id, length_L, entropy_band, question_variant) is unique per row.

SoccerNet rows

The 78 SoccerNet-derived rows (event class red_card) appear in natural_video/questions.parquet but the underlying mp4s are not redistributed here — the SoccerNet source is NDA-gated. Use the source_provenance JSON column to pull the originals from https://www.soccer-net.org/data after signing the SoccerNet NDA. See LICENSES/SoccerNet-NOTE.txt.

Licenses

split / source	license	per-row tag	file
Synthetic streams (text, synthetic_video)	CC BY 4.0	CC-BY-4.0	LICENSES/synthetic-and-easyhuman.txt
EasyHuman (easyhuman config)	CC BY 4.0	CC-BY-4.0	LICENSES/synthetic-and-easyhuman.txt
EPIC-Kitchens-100 derived clips (natural_video, source_dataset=epic-kitchens-100)	CC BY-NC 4.0 (research use only)	CC-BY-NC-4.0	LICENSES/EPIC-Kitchens-100-CC-BY-NC-4.0.txt
SoccerNet derived rows (natural_video, source_dataset=soccernet; provenance only)	NDA-gated source; not redistributed	SoccerNet-NDA	LICENSES/SoccerNet-NOTE.txt

The top-level license: cc-by-4.0 tag on this card refers to the Anonymous-Authors-owned splits (synthetic + EasyHuman) and to this card, metadata, and code only. EPIC-Kitchens-100 derivatives remain CC BY-NC 4.0.

Citation

Anonymous, "Substream Recollection," 2026 (anonymized for review).