Meow-10K: Feline Multimodal Training Dataset

• Meow-10K is a multisource feline training dataset comprising 10,831 samples across video, audio, time-series, and text modalities.
• It enables robust feline intent inference under semantic aliasing by integrating biometric data with natural-language supervision.
• The dataset supports Stage 2 multimodal fine-tuning in Meow-Omni 1 while remaining distinct from the expert-verified MeowBench evaluation suite.

Meow-10K is a multisource training dataset introduced with the quad-modal multimodal LLM Meow-Omni 1. It is defined as a dataset of 10,831 feline samples spanning varying modality combinations—video, audio, biometrics, and text-derived supervision—and is used for Stage 2 multimodal specialization rather than for held-out evaluation (Hu et al., 9 May 2026). Within the Meow-Omni 1 framework, its central purpose is to support feline intent inference under semantic aliasing, the setting in which identical observable signals can correspond to different latent states depending on physiological context. The paper distinguishes Meow-10K sharply from MeowBench, which is the expert-verified evaluation benchmark; Meow-10K trains, whereas MeowBench evaluates (Hu et al., 9 May 2026).

1. Definition and dataset identity

Meow-10K is presented as a diverse, multi-source dataset of 10,831 feline samples with varying modality combinations and natural-language descriptions (Hu et al., 9 May 2026). The dataset belongs to the released Meow-Omni 1 pipeline, which the paper describes as including model weights, training framework, and the Meow-10K dataset (Hu et al., 9 May 2026). Its role is therefore infrastructural as well as experimental: it is the principal multimodal fine-tuning corpus through which a general backbone is specialized to feline ethology.

The dataset is explicitly not the benchmark. The paper states that Meow-Omni 1 is fine-tuned using Meow-10K, whereas MeowBench is a separate held-out evaluation suite for measuring multimodal reasoning and intent recognition (Hu et al., 9 May 2026). This distinction is central because both resources belong to the same project and both involve multimodal intent-oriented supervision, but they occupy different positions in the training and evaluation pipeline.

Each Meow-10K sample consists of a natural-language query-response pair and, where applicable, an intention label drawn from a unified 30-class taxonomy (Hu et al., 9 May 2026). The dataset is therefore not merely a raw archive of media. It is structured as a supervised, instruction-style corpus designed for multimodal sequence modeling.

2. Scientific motivation and task formulation

The dataset is motivated by the paper’s claim that feline behavior exhibits semantic aliasing: identical external signals, such as a purr, may correspond to distinct internal states depending on physiological context (Hu et al., 9 May 2026). Meow-10K is the training resource intended to operationalize this claim by exposing the model to combinations of external behavior streams and biometric time-series.

In the paper’s framing, intent inference is not reducible to visual or acoustic classification. The authors define the problem around latent intention variables and explicitly argue that grounding the model in biometric data enables disambiguation of outwardly similar behaviors (Hu et al., 9 May 2026). This suggests that Meow-10K should be understood less as a conventional action-recognition dataset and more as a corpus for multimodal latent-state inference.

The dataset’s use of a unified 30-class intention taxonomy reinforces this orientation. Although examples listed in the appendix include behavior-like classes such as Feed, Groom, Rest, Run, Walk, active_climbing, inactive_lying.resting, maintenance_scratching, maintenance_shake.head, and other_social.allogrooming, the paper consistently describes the taxonomy as one of intention rather than merely overt motion categories (Hu et al., 9 May 2026). Because the exact reconciled taxonomy is not cleanly tabulated in the provided text, the safest statement is that the paper repeatedly specifies a 30-class intent schema.

3. Composition, modalities, and internal sources

Meow-10K is assembled from multiple data pipelines rather than a single synchronized recording campaign (Hu et al., 9 May 2026). The paper identifies several source components:

• TS-only samples derived from two peer-reviewed accelerometer datasets
• Video-only clips from the Bain dataset
• Naturally synchronized audio-video pairs from AudioSet
• Standalone audio captions from Freesound and cat_class
• A small set of expert-verified synthetic quad-modal samples (Hu et al., 9 May 2026)

A key quantitative detail is that the abundant time-series pool was deliberately downsampled. The paper states that, to prevent TS-only data from overwhelming other modalities, it randomly subsamples 2,000 sequences from a larger pool of 383,853 labeled TS samples, and that after balancing modality representation the final Meow-10K dataset contains 10,831 training samples (Hu et al., 9 May 2026). No exact per-source counts are given for the non-TS subsets.

The dataset is explicitly heterogeneous in modality availability. The paper states that it contains samples with combinations such as A/V/TS, A/V, V/TS, A/TS, and single-modality inputs (Hu et al., 9 May 2026). Missing modalities are omitted rather than represented with placeholder tokens during model input construction. This design couples the dataset to a missing-modality training regime rather than to a fully observed multimodal setting.

The following table summarizes the source structure described in the paper.

Component	Source or construction	Reported supervision form
TS-only	Two accelerometer datasets; 2,000 subsampled sequences	Query-response pairs; intention labels where applicable
Video-only	Bain dataset	Action captions mapped to 30 intention classes
Audio-video	AudioSet synchronized clips	Audio-focused caption plus 30-class probability distribution
Audio-only	Freesound and cat_class	Descriptive captioning, no pre-assigned intention label
Synthetic quad-modal	Intent-matched synthesis with expert verification	Query-response pairs and intent-matched multimodal supervision

This composition implies that Meow-10K is not annotation-homogeneous. Hard labels, soft label distributions, and caption-style responses coexist within the same training resource (Hu et al., 9 May 2026).

4. Physiological modality and preprocessing pipeline

Although Meow-Omni 1 is framed as a quad-modal system incorporating biometrics, the directly documented physiological channel in Meow-10K is triaxial accelerometry (Hu et al., 9 May 2026). The paper states that it adopts two accelerometer datasets from peer-reviewed studies and repeatedly refers to acceleration along the X, Y, and Z axes.

The time-series preprocessing pipeline is described with moderate specificity. The original accelerometer streams are sampled at 30 Hz or 60 Hz and are then aggregated into second-level signals by averaging the measurements within each second (Hu et al., 9 May 2026). These cleaned continuous streams are segmented into fixed-length windows of 5, 7, 10, and 15 seconds, and the prediction target is the behavioral label at a future offset of 1, 2, 3, or 5 seconds ahead (Hu et al., 9 May 2026). The segmentation is constrained so that windows do not cross individuals or discontinuities.

For video-only data, the paper describes a VLM-based event-localization pipeline that performs a coarse scan at approximately 1.5-second intervals, rescans gaps at approximately 0.5-second intervals, performs dense localization within a 4.0-second window at 0.20-second interval, and then extracts a final observation clip of 6.0 seconds (Hu et al., 9 May 2026). The exact asymmetrical temporal windowing formula is garbled in the provided text, but the fixed 6.0 s clip length is explicit.

For audio-video samples derived from AudioSet, temporal alignment is inherited from the video preprocessing stage, so the A/V subset is naturally synchronized internally (Hu et al., 9 May 2026). By contrast, the paper explicitly states that naturally synchronized quad-modal datasets do not exist in this setting; quad-modal training examples are synthesized by matching unimodal data that share the same intention label (Hu et al., 9 May 2026). This is an important limitation of the resource.

5. Annotation regimes and relation to MeowBench

Meow-10K uses different annotation strategies for different modalities (Hu et al., 9 May 2026). The paper describes these regimes explicitly.

For time-series, the labeling strategy is Next-Behaviour Prediction (NBP). Invalid or semantically ambiguous labels, such as “other” categories, are discarded, and transient intermediate movements that do not clearly represent stable intention are excluded (Hu et al., 9 May 2026). Query-response pairs are then generated for supervised instruction-style training.

For video-only clips, a VLM-based pipeline detects action onsets and produces natural-language action captions. Each caption is then mapped to one of the 30 intention classes by a secondary LLM, specifically Qwen3.5-35B, and this mapping is reviewed and validated by the same expert groups that curated MeowBench (Hu et al., 9 May 2026).

For audio-only clips, no pre-assigned intention label is given. Instead, the model is trained to produce a descriptive caption from the audio signal (Hu et al., 9 May 2026). This makes the audio-only subset supervision objective descriptively generative rather than directly classificatory.

For synchronized audio-video pairs, the paper describes an AST-based filtering pipeline with explicit thresholds: clips are retained if the cat-sound score is greater than 0.10, discarded if it is less than 0.03, and for scores in 0.03–0.10 a stationary noise reduction step with suppression factor 0.85 is applied, after which the clip is retained only if the score remains above 0.03 (Hu et al., 9 May 2026). The resulting A/V samples receive an audio-focused caption and a probability distribution over the 30 intention classes, which serves as a soft supervision target.

For synthetic quad-modal samples, synchronized A/V pairs are matched with TS samples sharing the same intention label and then verified by expert reviewers (Hu et al., 9 May 2026). The paper does not provide a separate expert-count statistic for Meow-10K itself, but it does state that the same professional feline ethologists involved in MeowBench review related mappings and synthesis steps.

The relation to MeowBench is therefore functional rather than definitional. MeowBench is the 527-sample expert-verified held-out benchmark used to report accuracy, while Meow-10K is the 10,831-sample training set used to specialize the model (Hu et al., 9 May 2026). Confusing the two would misstate the training and evaluation design.

6. Role in Meow-Omni 1 training, strengths, and limitations

Meow-10K is used in Stage 2: Multimodal Specialization of Meow-Omni 1 (Hu et al., 9 May 2026). Stage 1 is separate: it uses 383,853 labeled TS samples to train only the time-series projector, with the LLM and TS encoder frozen. In Stage 2, by contrast, the model is fine-tuned on Meow-10K, with all encoders and projectors frozen and only the LLM backbone updated (Hu et al., 9 May 2026). This two-stage curriculum indicates that Meow-10K is the dataset through which multimodal alignment is converted into downstream ethological competence.

The model input interface includes the TS control tokens <|ts_start|>, <|ts_unit|>, and <|ts_end|>, with the processor dynamically expanding the TS unit placeholder to match the number of hidden states produced by the encoder (Hu et al., 9 May 2026). TS embeddings are projected into the LLM hidden dimension and interleaved with visual and linguistic embeddings. The paper states that the model can process variable-length sequences containing any subset of the available modalities, and that missing modalities are absent from the input sequence (Hu et al., 9 May 2026). Meow-10K therefore directly supports a missing-modality instruction-tuning regime.

The conceptual objective is framed as intent inference over multimodal histories, while the practical Stage 2 training loss is described as next-token prediction loss over the full multimodal sequence, including special tokens (Hu et al., 9 May 2026). This dual description indicates that Meow-10K functions simultaneously as an instruction-tuning corpus and as the data substrate for latent-intent reasoning.

The paper’s main empirical result is reported on MeowBench rather than on Meow-10K directly: Meow-Omni 1 reaches 71.16% Top-1 accuracy on MeowBench and outperforms the listed vision-language and omni-modal baselines (Hu et al., 9 May 2026). Because Meow-10K is the Stage 2 specialization dataset behind this result, the performance claim is evidence for the utility of the training corpus, but not a direct standalone benchmark score for Meow-10K itself.

Several strengths of Meow-10K are explicit in the paper. It supports mixed-modality learning, includes natural-language supervision, introduces a unified 30-class intention taxonomy, and exposes the model to biometric time-series rather than restricting training to audio-visual cues (Hu et al., 9 May 2026). A plausible implication is that the dataset is designed less for narrow modality-specific classification than for robustness under realistic observational incompleteness.

The limitations are equally important. The paper does not provide the number of cats, sessions, or total hours; does not give per-source counts beyond the 2,000 TS subsample; does not report a formal inter-annotator agreement statistic; and does not describe richer physiological channels beyond accelerometry for Meow-10K (Hu et al., 9 May 2026). It also notes that naturally synchronized quad-modal datasets do not exist in this setting, so part of the multimodal supervision is necessarily synthetic. The paper explicitly states that the dataset remains inherently synthetic and that future verification should be conducted on entirely native, independently collected, isolated real-world data (Hu et al., 9 May 2026).

A final terminological caveat is that the string “Meow” is used elsewhere in the literature for other resources and projects. For example, V2Meow is a video-to-music generation system and does not introduce any dataset named Meow-10K (Su et al., 2023). Meow-10K therefore refers specifically to the feline ethology training dataset released with Meow-Omni 1, not to the broader set of unrelated “Meow” acronyms or model names appearing in other domains (Hu et al., 9 May 2026).

In summary, Meow-10K is a 10,831-sample multisource training dataset for multimodal feline intent modeling. It combines video, audio, accelerometer-based biometric time-series, and text supervision in an instruction-style format, is used for Stage 2 fine-tuning of Meow-Omni 1, and is methodologically distinguished from the smaller expert-verified MeowBench evaluation suite (Hu et al., 9 May 2026). Its principal significance lies in attempting to instantiate a physiologically grounded approach to computational ethology under missing-modality conditions, while its principal constraints are heterogeneous supervision, limited native synchronization, and incomplete reporting of composition metadata.