multivent-raw-features

Per-chunk features for the 143,288 short-video chunks in hltcoe/multivent-raw: ASR transcripts, OCR text, and a family of dense embeddings. This is a companion repo — the base media (videos/, keyframes/) and the retrieval/claim annotations/ live in multivent-raw. Every artifact here is per-chunk and joinable to that repo (and to the other artifacts here) by chunk_id.

---

At a glance

artifact	dir	size	one record per	vector dim
ASR (Qwen3-ASR-1.7B)	asr/qwen3asr1p7b/	1.3 GB	chunk	—
OCR (PaddleOCR-VL-1.5)	ocr/ppocrvl15/	5.0 GB	chunk	—
Vision emb (Qwen3-VL-Emb 2B)	embeddings/kf_uni5s-vizemb_qwen3vlemb2b/	31 GB	keyframe	2048
Vision emb (Qwen3-VL-Emb 8B)	embeddings/kf_uni5s-vizemb_qwen3vlemb8b/	61 GB	keyframe	4096
OCR-text emb (Qwen3-Emb 8B)	embeddings/kf_uni5s-ocr_ppocrvl15-txtemb_qwen3emb8b/	49 GB	keyframe	4096
OmniEmbed-01	embeddings/omniemb_omniembed01/	2.2 GB	chunk	3584
OmniEmbed-01 (mv)	embeddings/omniemb_omniembed01mv/	2.2 GB	chunk	3584
Omni-Nemotron-3B	embeddings/omniemb_omninemotron3b/	1.4 GB	chunk	2048
Video emb (Qwen3-VL-Emb 8B)	embeddings/videmb_qwen3vlemb8b/	2.5 GB	chunk	4096

Everything is packed as WebDataset shard_NNNNNN.tar (×667), sharded identically to multivent-raw (shard 42 here holds the same chunks as shard 42 there). ~155 GB total. All embeddings are float32 and L2-normalised (cosine similarity == inner product).

---

Directory layout

multivent-raw-features/
├── README.md
│
├── asr/
│   └── qwen3asr1p7b/                                        ← per-chunk ASR (Qwen3-ASR-1.7B)
│       ├── catalog.csv
│       └── shard_NNNNNN.tar   (×667)
│
├── ocr/
│   └── ppocrvl15/                                           ← per-frame OCR text (PaddleOCR-VL-1.5)
│       ├── catalog.csv
│       └── shard_NNNNNN.tar   (×667)
│
└── embeddings/
    ├── kf_uni5s-vizemb_qwen3vlemb2b/                        ← per-keyframe vision emb, dim 2048
    ├── kf_uni5s-vizemb_qwen3vlemb8b/                        ← per-keyframe vision emb, dim 4096
    ├── kf_uni5s-ocr_ppocrvl15-txtemb_qwen3emb8b/            ← per-keyframe text emb of ppocrvl15 OCR, dim 4096
    │   ├── catalog.csv
    │   └── shard_NNNNNN.tar   (×667)
    │
    ├── omniemb_omniembed01/                                 ← per-chunk emb (OmniEmbed-01),      dim 3584
    ├── omniemb_omniembed01mv/                               ← per-chunk emb (OmniEmbed-01 mv),   dim 3584
    ├── omniemb_omninemotron3b/                              ← per-chunk emb (Omni-Nemotron-3B),  dim 2048
    └── videmb_qwen3vlemb8b/                                 ← per-chunk video emb (Qwen3-VL-Emb 8B), dim 4096
        └── shard_NNNNNN.tar   (×667)                        ← (no catalog.csv — see Catalogs)

The three per-keyframe embeddings and the ASR/OCR dirs each ship a catalog.csv. The four chunk-level embeddings (omniemb_*, videmb_*) ship shards only — no catalog (see Catalogs).

---

Identifiers

Identical scheme to multivent-raw.

field	example	what it identifies
chunk_id	XM5xOIzL_vSkGAKR_0000	one chunk; the join key across all artifacts
video_id	XM5xOIzL_vSkGAKR	the source video the chunk came from
frame tNNNNNN	t000005	a keyframe within a chunk, at second NNNNNN of the chunk

• chunk_id == f"{video_id}_{chunk_index:04d}" (always 4-digit padded).
• Keyframes are sampled every 5 s, so frame ids are t000000, t000005, ….
• No id starts with -, so filenames are safe for tar/find/xargs.

To resolve a frame's timestamp in the source video, or to fetch the keyframe .jpg / source .mp4, join to keyframes/ and videos/ in the multivent-raw repo (same chunk_id, same shard_index).

---

In-shard file names

Inside every shard, members follow <chunk_id>.<artifact_tag>.<extension>. The tag is the artifact directory name with - → .:

artifact directory	tag	member example
asr/qwen3asr1p7b/	asr_qwen3asr1p7b	<cid>.asr_qwen3asr1p7b.json
ocr/ppocrvl15/	kf_uni5s.ocr_ppocrvl15	<cid>.kf_uni5s.ocr_ppocrvl15.jsonl
embeddings/kf_uni5s-vizemb_qwen3vlemb2b/	kf_uni5s.vizemb_qwen3vlemb2b	<cid>.kf_uni5s.vizemb_qwen3vlemb2b.npz
embeddings/kf_uni5s-vizemb_qwen3vlemb8b/	kf_uni5s.vizemb_qwen3vlemb8b	<cid>.kf_uni5s.vizemb_qwen3vlemb8b.npz
embeddings/kf_uni5s-ocr_ppocrvl15-txtemb_qwen3emb8b/	kf_uni5s.ocr_ppocrvl15.txtemb_qwen3emb8b	<cid>.kf_uni5s.ocr_ppocrvl15.txtemb_qwen3emb8b.npz
embeddings/omniemb_omniembed01/	omniemb_omniembed01	<cid>.omniemb_omniembed01.npz
embeddings/omniemb_omniembed01mv/	omniemb_omniembed01mv	<cid>.omniemb_omniembed01mv.npz
embeddings/omniemb_omninemotron3b/	omniemb_omninemotron3b	<cid>.omniemb_omninemotron3b.npz
embeddings/videmb_qwen3vlemb8b/	videmb_qwen3vlemb8b	<cid>.videmb_qwen3vlemb8b.npz

The stem before the first . is always the chunk_id — WebDataset uses it to group members for the same chunk into one sample.

---

Per-artifact details

asr/qwen3asr1p7b/

Per chunk: one .asr_qwen3asr1p7b.json with language detection, VAD, diarization, per-speaker voice embeddings, and a segment-level transcript (abridged):

{
  "chunk_id":    "XM5xOIzL_vSkGAKR_0000",
  "video_id":    "XM5xOIzL_vSkGAKR",
  "chunk_index": 0,
  "duration_s":  38.824,
  "language":    {"hint": null, "detected": "Russian", "from_dir": false},
  "asr_backend": "qwen",
  "vad":         [{"start": 1.19, "end": 4.05}, ...],
  "diarization": [{"start": 1.19, "end": 4.05, "speaker_id": "SPEAKER_00"}, ...],
  "overlap":     [...],
  "embeddings":  [{"speaker_id": "SPEAKER_00", "vector": [...], "n_turns": 3}, ...],
  "transcript": {
    "timestamp_resolution": "segment",
    "segments": [{"start": 1.19, "end": 4.05, "text": "...", "speaker_id": "SPEAKER_00", "flags": []}, ...],
    "words":    []
  },
  "qc_flags": []
}

The transcript text is the concatenation of transcript["segments"][*]["text"]. The per-speaker embeddings here are voice embeddings from diarization — unrelated to the embeddings/ artifact dirs. Only chunks with audio produce a record.

catalog.csv (138,328 rows): chunk_id, video_id, chunk_index, shard_index, duration_s, language_detected, n_speakers, n_segments, n_words, n_qc_flags, vad_coverage_sec.

ocr/ppocrvl15/

Per chunk: one .kf_uni5s.ocr_ppocrvl15.jsonl, one JSON object per keyframe (in tNNNNNN order, length == the chunk's frame_count):

{
  "frame":   "t000000",
  "raw":     "...model output with <|LOC_NNN|> coordinate tokens...",
  "cleaned": "...repetition-loop artifacts trimmed; LOC tokens preserved...",
  "txt":     "...LOC tokens stripped, whitespace tidied — ready for grep / text embedders..."
}

Most consumers want txt. cleaned keeps the spatial <|LOC_N|> layout tokens (coordinate index 0–999); raw is verbatim model output.

catalog.csv: chunk_id, video_id, chunk_index, shard_index, n_frames, n_frames_repetition_cleaned.

embeddings/ — per-keyframe (vision & OCR-text)

kf_uni5s-vizemb_qwen3vlemb2b, kf_uni5s-vizemb_qwen3vlemb8b, and kf_uni5s-ocr_ppocrvl15-txtemb_qwen3emb8b. Per chunk: one .npz with one row per keyframe.

key	shape	dtype
keyframe_ids	(N,)	<U7 (t000000, …)
embeddings	(N, D)	float32, L2-normalised

• vision (vizemb_qwen3vlemb2b / qwen3vlemb8b): keyframes encoded by Qwen3-VL-Embedding; D = 2048 (2B) / 4096 (8B). N == frame_count.
• OCR-text (ocr_ppocrvl15-txtemb_qwen3emb8b): each keyframe's OCR txt encoded by Qwen3-Embedding-8B; D = 4096. M ≤ N — frames whose OCR txt was empty are skipped (≈11 % of frames for ppocrvl15); chunks with no text in any frame produce no member. Row i describes frame keyframe_ids[i].

catalog.csv: vision — chunk_id, video_id, chunk_index, shard_index, n_frames, dim; OCR-text — …, shard_index, n_frames_embedded, n_frames_skipped, dim.

embeddings/ — chunk-level (whole-chunk vectors)

omniemb_omniembed01, omniemb_omniembed01mv, omniemb_omninemotron3b, and videmb_qwen3vlemb8b. Per chunk: one .npz with exactly one row for the whole chunk — keyframe_ids holds the chunk_id, not a frame id:

key	shape	dtype
keyframe_ids	(1,)	<U21 (the chunk_id)
embeddings	(1, D)	float32, L2-normalised

dir	model	D
omniemb_omniembed01	OmniEmbed-01	3584
omniemb_omniembed01mv	OmniEmbed-01 (mv variant)	3584
omniemb_omninemotron3b	Omni-Nemotron-3B	2048
videmb_qwen3vlemb8b	Qwen3-VL-Embedding-8B (whole chunk)	4096

These four ship without a catalog.csv (see below).

---

Catalogs

Each ASR/OCR dir and each per-keyframe embedding dir has a catalog.csv sharing the prefix (chunk_id, video_id, chunk_index, shard_index), so any pair joins on chunk_id:

import pandas as pd
ocr    = pd.read_csv("ocr/ppocrvl15/catalog.csv")
txtemb = pd.read_csv("embeddings/kf_uni5s-ocr_ppocrvl15-txtemb_qwen3emb8b/catalog.csv")
df = ocr.merge(txtemb, on="chunk_id", suffixes=("_ocr", "_txtemb"))

The four chunk-level embeddings (omniemb_*, videmb_*) have no catalog. To list their chunks, either enumerate .npz stems from the tars, or borrow any other artifact's catalog — the chunk_id/shard_index mapping is identical across every artifact (here and in multivent-raw). For the full chunk universe use multivent-raw's videos/catalog.csv (143,288 rows).

---

Loading examples

One chunk's embedding, by name

import io, tarfile, numpy as np

CID, SHARD = "XM5xOIzL_vSkGAKR_0000", 0
TAR = f"embeddings/kf_uni5s-vizemb_qwen3vlemb8b/shard_{SHARD:06d}.tar"
MEMBER = f"{CID}.kf_uni5s.vizemb_qwen3vlemb8b.npz"
with tarfile.open(TAR) as tf:
    data = np.load(io.BytesIO(tf.extractfile(MEMBER).read()))
print(data["keyframe_ids"])      # ['t000000' 't000005' ...]
print(data["embeddings"].shape)  # (N, 4096), L2-normalised

# chunk-level artifact: keyframe_ids is [chunk_id], embeddings is (1, D)
TAR = f"embeddings/videmb_qwen3vlemb8b/shard_{SHARD:06d}.tar"
with tarfile.open(TAR) as tf:
    v = np.load(io.BytesIO(tf.extractfile(f"{CID}.videmb_qwen3vlemb8b.npz").read()))
print(v["embeddings"].shape)     # (1, 4096)

ASR / OCR from a shard

import tarfile, json
with tarfile.open("ocr/ppocrvl15/shard_000000.tar") as tf:
    for m in tf:
        if not m.name.endswith(".jsonl"):
            continue
        for line in tf.extractfile(m).read().decode().splitlines():
            rec = json.loads(line)
            print(rec["frame"], rec["txt"][:60])
        break

WebDataset (multi-artifact, joined by chunk_id)

import webdataset as wds
url = "{ocr/ppocrvl15,embeddings/kf_uni5s-ocr_ppocrvl15-txtemb_qwen3emb8b}/shard_000000.tar"
ds = wds.WebDataset(url, shardshuffle=False).decode()
for sample in ds:
    chunk_id = sample["__key__"]
    ocr      = sample.get("kf_uni5s.ocr_ppocrvl15.jsonl")
    txt_emb  = sample.get("kf_uni5s.ocr_ppocrvl15.txtemb_qwen3emb8b.npz")
    ...

---

Sharding

667 shards of ~210 chunks each. A chunk lives in exactly one shard, and the shard index matches multivent-raw — shard 42 of any artifact (here or there) describes the same set of chunks. Shards are independent, so wds.WebDataset(shardshuffle=True) gives IID-ish batches.

---

Relationship to multivent-raw

This repo is features only. For the source media and evaluation data, see hltcoe/multivent-raw: videos/ (.mp4 + chunk JSON), keyframes/uniform_5s/ (.jpg), keyframe-captions/, the other OCR/ASR/embedding backends, and annotations/ (queries, personas, topics, reference claims, qrels). Join by chunk_id.