Derived signals for the microvent core release: per-keyframe OCR text, per-chunk ASR transcripts, and an embedding zoo (keyframe-level vision, keyframe-OCR text, audio-level, video-level, omni-modal).

This card covers only the features. For the source videos, audio, keyframes, and the public eval annotations, see the microvent dataset card. All artifacts here key on the same chunk_id and follow the same WebDataset shard layout, so joining feature shards back to the core release is a straight tar-member lookup.

Directory layout

microvent-features/
├── README.md
│
├── ocr/
│   └── ppocrvl15/                                           ← per-frame OCR text (PaddleOCR-VL-1.5, cleaned)
│       ├── catalog.csv
│       └── shard_NNNNNN.tar   (×5)
│
├── asr/
│   └── qwen3asr1p7b/                                        ← per-chunk ASR (Qwen3-ASR-1.7B)
│       ├── catalog.csv
│       └── shard_NNNNNN.tar   (×5)
│
└── embeddings/                                              ← per-chunk .npz, keyed by chunk_id
    │
    │ ── vision over uniform-5s keyframes ──
    ├── kf_uni5s-vizemb_qwen3vlemb2b/                        ← Qwen3-VL-Embedding-2B,  dim 2048
    ├── kf_uni5s-vizemb_qwen3vlemb8b/                        ← Qwen3-VL-Embedding-8B,  dim 4096
    ├── kf_uni5s-vizemb_siglip2so400m512/                    ← SigLIP2-So400M/512,     dim 1152
    │
    │ ── text embedding of keyframe OCR ──
    ├── kf_uni5s-ocr_ppocrvl15-txtemb_qwen3emb8b/            ← Qwen3-Embedding-8B over ppocrvl15 text, dim 4096
    │
    │ ── audio-level (one vector(s) per chunk's audio) ──
    ├── audemb_glap/                                         ← GLAP,                    dim 1024
    ├── audemb_jinav5omnismall/                              ← Jina-v5-omni-small,      dim 1024
    ├── audemb_largerclapgeneral/                            ← Larger-CLAP-general,     dim 512
    ├── audemb_lcoomni7b/                                    ← LCO-Embedding-Omni-7B (audio), dim 3584
    ├── audemb_omniembed01/                                  ← OmniEmbed-v0.1 (audio),  dim 3584
    ├── audemb_omninemotron3b/                               ← Omni-Embed-Nemotron-3B (audio), dim 2048
    │
    │ ── video-level (one vector per chunk's full video) ──
    ├── videmb_lcoomni7b/                                    ← LCO-Embedding-Omni-7B (video), dim 3584
    ├── videmb_omninemotron3b/                               ← Omni-Embed-Nemotron-3B (video), dim 2048
    ├── videmb_qwen3vlemb8b/                                 ← Qwen3-VL-Embedding-8B,   dim 4096
    │
    │ ── omni-modal (joint audio+video per chunk) ──
    ├── omniemb_lcoomni7b/                                   ← LCO-Embedding-Omni-7B (omni), dim 3584
    ├── omniemb_omniembed01/                                 ← OmniEmbed-v0.1 (omni),   dim 3584
    └── omniemb_omninemotron3b/                              ← Omni-Embed-Nemotron-3B (omni), dim 2048

Model cards for everything listed above are linked from the embedding table further down.

Each artifact directory contains the same two-file pattern: a catalog.csv and the shard_NNNNNN.tar WebDataset shards. The newer embedding directories ship 3 shards (~~314 chunks each); the older keyframe-vision and OCR-text-embedding directories, plus ocr/ and asr/, ship 5 shards (~~189 chunks each) matching the core release. Some newer embedding directories may be missing their catalog.csv pending a backfill; the chunk membership is always recoverable from the tar TOC in that case.

Identifiers and join keys

Same chunk_id / video_id / tNNNNNN scheme as the core release. The filename of every tar member starts with the chunk_id of the source chunk, so a WebDataset loader will group features and core artifacts into the same sample automatically when you wds.WebDataset(...) over both shard sets.

In-shard file names

<chunk_id>.<artifact_tag>.<extension>

artifact directory	tag	member
`ocr/ppocrvl15/`	`kf_uni5s.ocr_ppocrvl15`	`.jsonl` (one line per frame)
`asr/qwen3asr1p7b/`	`asr_qwen3asr1p7b`	`.json`
`embeddings/kf_uni5s-vizemb_qwen3vlemb2b/`	`kf_uni5s.vizemb_qwen3vlemb2b`	`.npz`
`embeddings/kf_uni5s-vizemb_qwen3vlemb8b/`	`kf_uni5s.vizemb_qwen3vlemb8b`	`.npz`
`embeddings/kf_uni5s-vizemb_siglip2so400m512/`	`kf_uni5s.vizemb_siglip2so400m512`	`.npz`
`embeddings/kf_uni5s-ocr_ppocrvl15-txtemb_qwen3emb8b/`	`kf_uni5s.ocr_ppocrvl15.txtemb_qwen3emb8b`	`.npz`
`embeddings/audemb_*/`	`audemb_<model>`	`.npz`
`embeddings/videmb_*/`	`videmb_<model>`	`.npz`
`embeddings/omniemb_*/`	`omniemb_<model>`	`.npz`

The stem before the first . is always the chunk_id.

Per-artifact details

OCR (`ocr/ppocrvl15/`)

PaddleOCR-VL-1.5 run per keyframe, then lightly cleaned. Each chunk contributes one <chunk_id>.kf_uni5s.ocr_ppocrvl15.jsonl file whose lines are one frame each, in tNNNNNN order. Each line is a JSON object with these fields:

field	type	meaning
`frame`	str	the `tNNNNNN` second-offset label for the keyframe
`raw`	str	the model's raw output, with bounding-box location tokens like `<
`cleaned`	str	the same string after light post-processing (the cleanup is conservative; for many frames `cleaned == raw`)
`txt`	str	the recognized text only, with all `<

ASR (`asr/qwen3asr1p7b/`)

Qwen3-ASR-1.7B run per chunk on the audio track. Each chunk contributes one <chunk_id>.asr_qwen3asr1p7b.json with whole-chunk transcript text plus per-segment timings. Chunks with has_audio=False (10 of 943) have no JSON member.

Embeddings (`embeddings/`)

Every .npz has the same two-array schema regardless of model or modality:

key	shape	dtype	meaning
`keyframe_ids`	`(N,)`	`<U*`	row labels: `tNNNNNN` for keyframe-level embeddings, `<chunk_id>` for chunk-level
`embeddings`	`(N, D)`	float32	one row per `keyframe_ids` entry; `D` is the model's output dim

Keyframe-level (kf_uni5s-...): N == frame_count from the keyframe catalog; keyframe_ids are the tNNNNNN strings. One row per keyframe.
Chunk-level (audemb_*, videmb_*, omniemb_*): N == 1 for most backends; keyframe_ids carries the chunk_id. A couple of audio backends segment internally and emit one row per internal window instead (audemb_glap and audemb_largerclapgeneral); for those, keyframe_ids carries window labels.

Embedding dims and model cards:

family	dir tag	dim	model card
vision (kf)	`vizemb_qwen3vlemb2b`	2048	Qwen/Qwen3-VL-Embedding-2B
vision (kf)	`vizemb_qwen3vlemb8b`	4096	Qwen/Qwen3-VL-Embedding-8B
vision (kf)	`vizemb_siglip2so400m512`	1152	google/siglip2-so400m-patch16-512
text (kf)	`txtemb_qwen3emb8b` over ppocrvl15	4096	Qwen/Qwen3-Embedding-8B
audio	`audemb_glap`	1024	mispeech/GLAP
audio	`audemb_jinav5omnismall`	1024	jinaai/jina-embeddings-v5-omni-small
audio	`audemb_largerclapgeneral`	512	laion/larger_clap_general
audio	`audemb_lcoomni7b`	3584	LCO-Embedding/LCO-Embedding-Omni-7B
audio	`audemb_omniembed01`	3584	Tevatron/OmniEmbed-v0.1
audio	`audemb_omninemotron3b`	2048	nvidia/omni-embed-nemotron-3b
video	`videmb_qwen3vlemb8b`	4096	Qwen/Qwen3-VL-Embedding-8B
video	`videmb_lcoomni7b`	3584	LCO-Embedding/LCO-Embedding-Omni-7B
video	`videmb_omninemotron3b`	2048	nvidia/omni-embed-nemotron-3b
omni	`omniemb_lcoomni7b`	3584	LCO-Embedding/LCO-Embedding-Omni-7B
omni	`omniemb_omniembed01`	3584	Tevatron/OmniEmbed-v0.1
omni	`omniemb_omninemotron3b`	2048	nvidia/omni-embed-nemotron-3b

Catalog columns (where the file exists):

chunk_id, shard_index, input_shard, source_member,
video_id, chunk_index, embedding_dim, embedding_rows, artifact_id

Sharding and joins

Chunk → shard assignment for ocr/, asr/, and the older keyframe-vision / OCR-text-embedding directories matches the core microvent release (5 shards). Newer embedding directories were processed in a different pass with 3 shards; the chunk-membership union is still the same 943 chunks, but the shard index will differ. If you need a single chunk-keyed table across everything, join on chunk_id (not on shard_index).

Pulling the data locally

Mirror the whole feature release or any subset with the hf CLI:

# everything
hf download hltcoe/microvent-features --repo-type dataset --local-dir ./microvent-features

# just OCR + ASR (skip the embedding zoo)
hf download hltcoe/microvent-features --repo-type dataset --local-dir ./microvent-features \
    --include "ocr/*" "asr/*"

# one specific embedding config
hf download hltcoe/microvent-features --repo-type dataset --local-dir ./microvent-features \
    --include "embeddings/kf_uni5s-vizemb_qwen3vlemb8b/*"

--local-dir writes plain files (no blob/symlink indirection); drop it to land in the standard ~/.cache/huggingface/hub/ layout instead.

Loading with `datasets`

Every feature directory is exposed as a separate config (so you only pay to download what you need):

import datasets
ocr  = datasets.load_dataset("hltcoe/microvent-features", "ocr_ppocrvl15",     split="train", streaming=True)
asr  = datasets.load_dataset("hltcoe/microvent-features", "asr_qwen3asr1p7b",  split="train", streaming=True)
viz  = datasets.load_dataset("hltcoe/microvent-features", "emb_kf_uni5s_vizemb_qwen3vlemb8b", split="train", streaming=True)

To join features with the core artifacts, point webdataset at both shard sets and let the chunk-id stem do the grouping:

import webdataset as wds
ds = wds.WebDataset([
    "videos/shard_{000000..000004}.tar",
    "embeddings/kf_uni5s-vizemb_qwen3vlemb8b/shard_{000000..000002}.tar",
]).decode()

License

HLTCOE-authored content (this README, the catalog.csv files, and all of the OCR / ASR / embedding outputs produced by HLTCOE-run pipelines) is released under Apache-2.0.
The upstream models used to generate these features (PaddleOCR-VL-1.5, Qwen3-ASR-1.7B, Qwen3-VL-Embedding-2B/8B, Qwen3-Embedding-8B, SigLIP2, GLAP, Jina-v5-omni-small, laion CLAP, Tevatron OmniEmbed, nvidia omni-embed-nemotron, LCO-Embedding-Omni) carry their own licenses; consult each model's card (linked in the embeddings table above) before redistributing the embedding vectors in a commercial setting.
The source video, audio, and keyframe content that these features describe lives in the microvent core release and is copyrighted by its respective original owners. Distributing these features alongside the source media is research / fair-use only.

Downloads last month: 5

Total file size:

763 MB

You need to agree to share your contact information to access this dataset

microvent-features