Lostgreen/EventLens-Train

Quick start · Training views · Annotations & media

EventLens provides event-based video supervision for boundaries, progression, and relations, alongside temporal grounding and video question answering. This release brings together frozen no-CoT training views, original L1/L2/L3 hierarchy annotations, and the shared JPEG frames used by those training views.

Training examples¹	Monitor examples	Original annotation files	Shared frames
37,428	1,950	11,061	2,921,513

¹ Counts refer to the default paper_mt_opd view. Other configurations reuse examples and media; their counts are not additive.

Video domains

Source-video domain distribution reproduced from the paper appendix. The figure describes videos used to construct the recovery tasks.

Task families

Task	What the model learns	Training examples	Config
Boundary	Locate event boundaries at different hierarchy levels	10,000	seg
Progression	Recognize the temporal progression of actions	10,000	prog
Relation	Recover event order and relationships between clips	10,000	relation
Temporal grounding	Locate a text-described moment in a video	4,791	tg
Video QA	Answer multiple-choice questions about video content	2,637	mcq

A glimpse of the data

These are three actual records from the frozen release. Answers retain their original format.

Family	Source video ID	Model input	Ground truth
Boundary	ZccajuCybns	One input with 61 frames	Six temporal intervals, wrapped in <events> tags
Progression	UAR9gmFr_m8	Two inputs with 4 frames each	B
Relation	2rkEYETtXH8	Three inputs with 2 frames each	213

Quick start

from datasets import load_dataset

REPO = "Lostgreen/EventLens-Train"
REVISION = "main"  # Use a commit SHA to pin an experiment.

train = load_dataset(REPO, "paper_mt_opd", split="train", revision=REVISION)
monitor = load_dataset(REPO, "paper_mt_opd", split="monitor_val", revision=REVISION)

annotations = load_dataset(
    REPO, "annotations_hier_v1", split="annotations", revision=REVISION
)

This loads records and media references. To train with the JPEGs, download and unpack the shared frame shards described below.

Training views

Choose paper_mt_opd for the full mix, a specialist view for teacher training, or a task config for an individual family.

Config	Contents	Train	Monitor
paper_mt_opd	Grounding + QA + Boundary + Progression + Relation	37,428	1,950
paper_seg_teacher	Grounding + QA + Boundary	17,428	1,350
paper_prog_teacher	Grounding + QA + Progression	17,428	1,500
paper_relation_teacher	Grounding + QA + Relation	17,428	1,500
tg	Temporal grounding	4,791	600
mcq	Video QA	2,637	600
seg	Boundary	10,000	150
prog	Progression	10,000	300
relation	Relation	10,000	300

Split meaning: monitor_val preserves the historical training monitor and overlaps with training. It is not an independent validation or test split. The repository stores it under val/.

Annotations & media

The repository separates training records, original annotations, and shared frames. Train, monitor, and specialist views all reference the same media files.

EventLens-Train/
├── train/<config>/        Parquet + training.jsonl.gz
├── val/<config>/          Parquet + training.jsonl.gz
├── annotations/
│   ├── hier_v1/           Parquet + original_json.tar.gz
│   └── hier_legacy/       Parquet + original_json.tar.gz
├── media/shards/          37 shared frame TARs
├── metadata/              Video links, frame indexes, hashes & audits
└── tools/                 Training converter + schema

Unpacking media/shards/*.tar at the repository root creates media/frames/.

Original hierarchy annotations

Config	Split	JSON files	Source composition
annotations_hier_v1	annotations	9,738	ET-Instruct
annotations_hier_legacy	annotations	1,323	792 ET-Instruct + 531 TimeLens

The collections preserve historical annotation versions. Labels, time intervals, and descriptions match the original JSONs; machine-specific video_path, source_video_path, and frame_dir values are replaced with portable provenance references. Parquet stores the original hierarchy in payload_json; the archives retain the original JSON filenames.

Join at the video level: use source_dataset + source_video_key from metadata/video_links.jsonl. Each of the 11,061 annotation IDs has one unambiguous source-video owner. Four readable clip IDs refer to multiple source paths, so full source keys are required to disambiguate them.

Annotation source	Unique annotated videos	Have training frames	Outside training media
ET-Instruct	10,035	9,108	927
TimeLens	531	1	530

All 30,750 hierarchy-derived train/monitor examples link to original hierarchy annotations. TG/QA records retain their upstream supervision and source-video provenance; they do not all have EventLens hierarchy labels. Original annotations also include videos not used by the training views.

Shared frame media

The 2,921,513 referenced JPEGs are packaged into 37 TAR shards, with 71.07 GiB of JPEG payload before archive overhead. Frame indexes record the source-video key, TAR member, byte size, and SHA256; the shard manifest records each TAR's SHA256.

from huggingface_hub import snapshot_download

snapshot_download(
    REPO, repo_type="dataset", revision=REVISION, local_dir="EventLens-Train"
)

DATASET_ROOT=/path/to/EventLens-Train
for shard in "$DATASET_ROOT"/media/shards/*.tar; do
  tar -xf "$shard" -C "$DATASET_ROOT"
done

Training inputs preserve frozen frame sampling, multiple-clip boundaries, and frame order, including reordered action blocks. The bundled media are sampled training frames; native video binaries and annotation-process frame folders are referenced through provenance and are not bundled. Optional source/presentation timestamps remain null, and cache frame indices are not native-video frame numbers.

Use with the training code

Each training view includes a portable runtime JSONL alongside its typed Parquet:

gzip -dc "$DATASET_ROOT/train/paper_mt_opd/training.jsonl.gz" > train.jsonl
gzip -dc "$DATASET_ROOT/val/paper_mt_opd/training.jsonl.gz" > val.jsonl

Set EasyR1's data.image_dir to DATASET_ROOT so relative nested frame lists resolve to the extracted media. Alternatively, convert a Hub split with the included script:

python "$DATASET_ROOT/tools/convert_to_training_jsonl.py" \
  --input Lostgreen/EventLens-Train --config paper_mt_opd --split train \
  --prompt-mode no_cot --media-root "$DATASET_ROOT" --check-media \
  --output train.jsonl

Pin --revision to the same commit used for the media download when reproducing an experiment. The conversion preserves prompt, messages, answer, problem_type, data_type, video grouping, and frame order.

Fields at a glance

Field	Meaning
sample_id / row_id	Stable content identity / occurrence within a frozen view
task_family / problem_type	Human-readable task family / exact training and reward routing
prompt / messages / answer	Model input and original ground truth
media[].frame_keys	Ordered relative JPEG paths for each model input
media[].source_video_id	Readable upstream video identifier
annotation_id	Primary hierarchy annotation link where available
training_metadata_json	Portable original timing, construction, and reward metadata

See the typed schema for the complete field definitions. sources/ and provenance/ are source references rather than downloadable payload directories.

Verification & provenance

The frozen default view was audited across all train/monitor references: every referenced JPEG exists, frame cache directories have consistent video provenance, and every hierarchy-derived row links to original annotations. All 11,061 published annotation JSONs were compared with their originals with zero semantic changes.

A 512-record conversion preflight covered all 12 problem types, and all 9 training plus 2 annotation configs passed standard load_dataset checks. Detailed coverage and content checks are included in metadata/media_coverage.json and metadata/annotation_content_verification.json; media hashes are in metadata/media_shards.json and metadata/media_index/.

This is a frozen release. Upstream source revisions were not recorded; those values remain unknown rather than being inferred.

Sources & license

Noncommercial use. EventLens-derived hierarchy annotations and proxy labels use CC-BY-NC-SA-4.0. Source-specific terms continue to apply to third-party records and media; see LICENSES.md.

Upstream source	Link	Terms
ET-Instruct-164K	Dataset	CC-BY-NC-SA-4.0
TimeLens-100K	Dataset	Custom academic/noncommercial terms, including the upstream EU restriction
TimeRFT / TimeR1	Project	Upstream and constituent dataset terms
LLaVA-Video-178K	Dataset	Upstream and constituent media terms

The code license is separate from the data and media terms.