Talk2NatureZebra finches / Multimodal caller-attribution dataset audit
V33DA: a zebra-finch caller benchmark, not a translation model
Dataset-card revision and code sample reviewed 2026 · Note reviewed 2026-10-23
V33DA is a caller-attribution task: given an already-detected zebra-finch vocalization and the birds visible at that moment, identify the bird that produced it. The pinned card reports 33,625 verified events from 10 physically distinct birds across three experiments, with synchronized audio, video and pose; caller labels derive from on-body accelerometer vibration. In the experiment-held-out split, all 5,829 BP05 events from three callers (brown, purple, yellow) form the test set. That is a genuine unseen-bird group according to the card, but not a clean explanation of generalization because several cohort conditions shift together. This is an engineering benchmark candidate, not a communication-meaning result.
What we reviewed · Pinned dataset-card, code and sample metadata reviewed; media uninspected
Reviewed the pinned V33DA dataset card, code/sample README, split-preparation script, training entry point and read-only Git-tree inventory. Fetched the sample README.md, metadata.json and split_summary.json over verified TLS; computed in-memory SHA-256 values matched the pinned Git blob IDs. See the manifest for checksums. No text copies were retained. No Parquet, audio, accelerometer, video, calibration, executable code or model bytes were acquired or run. Maintainer metadata review only; no specialist review. A follow-up static import/tree check confirmed that scripts/train.py imports v33da.data modules absent from the pinned public repository tree; the split-preparation script materializes only session and held-out-experiment assignments. No code was run.
Not independently reviewed by an avian, bioacoustics or ML/statistical-evaluation specialist.
Study at a glance
- Question
- Among the birds visible at an already-detected, isolated call, which individual produced it?
- Label source
- Caller labels are derived from on-body accelerometer vibration; accelerometer data must not be a model input.
- Full-release scale
- 33,625 events, 10 birds, 3 experiments, 16 recording days and about 74 hours as reported by the pinned card.
- Repository sample inventory
- 100 audio WAVs, 100 accelerometer WAVs, 100 MP4 clips, 30 calibration files and 4 root files; 62,503,976 bytes total in pinned Git metadata.
- Reported generalization split
- Full-release experiment-held-out split: train BP01 12,881 + BP02 8,610; validation BP01 3,338 + BP02 2,967; test BP05 5,829 from three unseen physical callers (brown 56, purple 2,140, yellow 3,633). This bundles experiment/group/date/year/cage changes and is not the sample partition.
- Review state
- Metadata-only; three small text files fetched in memory and discarded; no Parquet, media, code or model run and no Talk2Nature benchmark result.
- Session-disjoint split
- BP01 train/validation/test 12,881/1,008/2,330; BP02 4,455/4,155/2,967; BP05 2,029/1,706/2,094. The same birds occur across sessions, so this is not unseen-bird generalization.
- Code snapshot check
- The trainer imports v33da.data.collate, dataset, features and splits, but the inspected public tree has no v33da/data directory. The split-preparation script writes date-based session and held-out-experiment labels only; the advertised LOO path remains unverified.
Reading location: Pinned Hugging Face README: Task, Modalities, Ground-truth labels, Evaluation regimes and Limitations. Pinned GitHub README and included metadata.json/split_summary.json. Three text-file integrity checks and path/size/Git-blob manifest are linked from source 357. Follow-up: pinned scripts/train.py import lines, scripts/prepare_v33da_splits.py split mappings, and the complete Git tree for the v33da package.
What this does — and doesn’t — establish
Caller identity is not behavior, addressee, receiver response, intent or meaning. The full release comes from one aviary and recording geometry and excludes overlapping calls; 10 physical birds, not 33,625 events, are the independent animals. The BP05 test also changes experiment, group, year/date and cage; treat its outcome only as a bundled shift test. The session-disjoint split retains the same birds. The pinned GitHub tree lacks the v33da/data package imported by scripts/train.py, including dataset, feature and split modules; the split-preparation script writes date-based session and held-out-experiment labels only. The README advertises leave-one-bird-out, but that path and the trainer cannot be independently reproduced from this snapshot. Three small text files have verified hashes, but the sample Parquet says 100 rows while split JSON reports full-release totals; no sample-specific split or Parquet join has been verified. BP05 metadata names brown, purple, yellow and red, while the card and caller-count table list three callers. Red has no event-count group; whether it is a visible non-caller candidate or stale metadata is unknown. Candidate-set cardinality and chance baseline are therefore unknown. No media or Parquet bytes were inspected. The separate V33DA++ companion adds overlap and padded context buckets but still uses the same ten birds and supplies no independently observed receiver behavior; see the companion note.
What we can build on
Use V33DA only if the question is caller attribution. A future baseline should compare candidate-count chance with a simple candidate-aware model, exclude accelerometer/oracle labels and stable identity shortcuts, and keep complete physical birds and shared sessions out of train/test groups. The BP05 experiment-held-out score is a bundled shift test. Do not call caller prediction translation. For overlap handling and longer context, see the separate V33DA++ review; its repeated bucket is not an independent cohort.
A practical starting point
Metadata-only original summary; not admitted to Talk2Nature training or the private database.
Available starting point: Pinned dataset-card and repository documentation, plus three in-memory text-file integrity checks and a public Git-tree inventory. No third-party raw media, Parquet, software or weights were acquired. A static import/tree check found the trainer's v33da.data.collate/dataset/features/splits imports absent from the pinned repository tree; no code was run.
Reuse limits: The dataset card and nested sample README declare CC BY 4.0; the code repository's MIT license applies to software. These are separate upstream declarations, not independent legal clearance. Review dependencies, any individual media terms and intended use before reuse.
Our proposed experiment: A later, separately approved experiment could compare a simple candidate-ranker against candidate-count chance on a reviewed unseen-bird split. The convenience sample is not a locked test, no baseline has been run, and no performance result is claimed.
See the related tools and reading paths →
The next reading task
Before any run, require a pinned code snapshot containing the imported v33da.data modules and an independently auditable leave-one-bird-out split. Then ask independent avian-bioacoustics and ML/statistical reviewers to assess candidate identity mapping, the BP05 color entry, sample-to-full-release joins, grouping and baseline design. Keep all source media metadata-only until rights and a bounded protocol are separately approved.
Go to the sources
- V33DA zebra-finch caller-attribution dataset card, pinned revision ↗
The dataset card describes 33,625 physically verified caller events from 10 zebra finches in three experiments, with synchronized audio, video, pose and telemetry. Its bounded task is identifying which visible bird made an already-detected call; it has no independent behavior or meaning labels. The associated manuscript is listed as under review. No dataset files, code or model were acquired or run.
- V33DA benchmark code and sample, pinned GitHub revision ↗
The repository provides split preparation and caller-attribution baselines. Its top-level LICENSE is MIT for software; the nested 100-event real-media sample README declares CC BY 4.0 separately. A pinned Git-tree inventory lists 334 sample files totaling 62,503,976 bytes. Follow-up source inspection found that scripts/train.py imports v33da.data modules absent from the pinned public repository tree; the advertised leave-one-bird-out path is not independently reproducible from this snapshot. Three small text metadata files were checksummed in memory; no media or Parquet bytes were acquired.
- V33DA++ overlapping-call and long-context companion dataset ↗
The V33DA++ dataset card describes reviewer-verified overlapping zebra-finch calls and padded context windows for source separation, vocal-activity detection and audio-visual synchronization. It uses the same recordings and ten birds as V33DA.
Original editorial note by Talk2Nature contributors. AI-assisted, source-grounded synthesis; not a peer review or an independent replication. Linked works retain their own rights.