Bioscan-Traits Dataset

• Bioscan-Traits is a dataset of 80.8K natural-language morphological trait annotations for 19.1K insect images derived from BIOSCAN-5M.
• The pipeline combines DINOv2 features, a sparse autoencoder, species-contrastive latent selection, and multimodal language prompting to localize and describe body parts.
• This resource enhances trait-aware ecological analysis, improves species classification, and addresses limitations in manual trait annotation.

Bioscan-Traits is a large, automatically constructed dataset of image-level morphological trait annotations for insects, built on BIOSCAN-5M and generated by a modular pipeline that combines DINOv2 features, a sparse autoencoder, species-contrastive latent selection, and multimodal language-model prompting. It contains about 80,800 natural-language trait descriptions attached to 19,100 individual insect specimen images, or about 4.2 traits per image, and was introduced to address the “trait bottleneck” in ecology and evolution by linking biological images to explicit, interpretable trait descriptions (Pahuja et al., 2 Apr 2026). The source images come from BIOSCAN-5M, a multimodal dataset of 5,150,850 arthropod specimens with images, DNA barcodes, taxonomy, geography, and size information (Gharaee et al., 2024).

1. Dataset scope within the BIOSCAN ecosystem

Bioscan-Traits sits on top of BIOSCAN-5M rather than replacing it. BIOSCAN-5M is a multimodal dataset containing 5,150,850 arthropod specimens, about 98% insects, with 5.15M microscope images, standardized COI DNA barcodes, Barcode Index Numbers, hierarchical taxonomy, geography, and size-related metadata (Gharaee et al., 2024). Bioscan-Traits uses the subset of BIOSCAN-5M with species-level labels, which is about 9.2% of the full corpus, because species labels are needed for the species-contrastive trait-selection stage (Pahuja et al., 2 Apr 2026).

The annotated portion of Bioscan-Traits covers 736 species across 417 genera. The organisms are arthropods imaged as pinned or voucher specimens under lab conditions, with reasonably standardized images that still exhibit pose variation, preservation artifacts, and background clutter. This combination is important because the pipeline is designed to localize fine-grained body parts despite standardized specimen imaging not being perfectly uniform (Pahuja et al., 2 Apr 2026).

Quantity	Value
Unique images	19.1K
Trait samples	80.8K
Average traits per image	about 4.2
Species coverage	736 species
Genus coverage	417 genera

This positioning makes Bioscan-Traits a trait-rich derivative resource built from a much larger multimodal biodiversity foundation dataset. A plausible implication is that Bioscan-Traits functions as an intermediate supervision layer between specimen imagery and broader ecological or taxonomic analyses.

2. Meaning and format of “morphological traits”

In Bioscan-Traits, “morphological traits” are visually discernible physical features of specific anatomical parts, phrased as short natural-language descriptions that a human taxonomist would recognize as legitimate morphology (Pahuja et al., 2 Apr 2026). The annotations are image-level in the sense that each trait is associated to an image, but they are also region-anchored: the pipeline first identifies a specific region, such as a bounding box around a wing or antenna, and then generates a description of that region.

The released annotations are free-form rather than ontology-bound. No fixed trait ontology or discrete categorical schema is imposed. Instead, traits are short natural-language descriptions such as “Wing: transparent, elongated, with visible veins.”, “[Leg]: thin, elongated, light brown, segmented.”, “Antenna: thin, elongated, segmented, dark color.”, and “Wing: brown, translucent, folded, with visible veins.” (Pahuja et al., 2 Apr 2026). In practice, these descriptions group naturally into body parts, shape, color and pattern, and texture or surface properties.

The format complements BIOSCAN-5M’s existing modalities. BIOSCAN-5M already provides image files, DNA sequences, Barcode Index Numbers, taxonomic fields from phylum to species, geographic coordinates, and size-related fields such as image_measurement_value, area_fraction, and scale_factor (Gharaee et al., 2024). Bioscan-Traits adds explicit morphological language to this multimodal record. This suggests that the dataset turns morphology from a largely latent property of images into a directly queryable text-based representation.

3. Annotation pipeline

The Bioscan-Traits pipeline proceeds in three conceptual stages: extracting dense visual features with DINOv2, training a sparse autoencoder to obtain monosemantic and spatially grounded latent units, and prompting a multimodal LLM on the localized regions (Pahuja et al., 2 Apr 2026).

For visual encoding, the backbone is DINOv2-base (ViT-B/14). The authors extract patch-token features from the penultimate ViT layer, specifically layer 10, flatten them into a dense vector $x \in \mathbb{R}^d$, and train a single-layer ReLU sparse autoencoder with a 32× expansion, i.e. hidden width 24,576: $$h = W_e (x - b_d) + b_e,\quad g(x) = \text{ReLU}(h),\quad \tilde{x} = W_d\, g(x) + b_d,$$ with objective

$$\mathcal{J}(\phi) = \| x - \tilde{x} \|_2^2 + \alpha\, \mathcal{R}(g(x)).$$

The reported hyperparameters include sparsity coefficient $$\alpha \in \{2\times 10^{-4}, 4\times 10^{-4}, 8\times 10^{-4}\}$$, learning rate $5\times 10^{-4}$ or $1\times 10^{-3}$, batch size 16,384, and features from ViT layer 10 (Pahuja et al., 2 Apr 2026).

A central empirical claim is that many latent units become monosemantic and spatially grounded. Visual inspection showed examples in which neuron 4852 activates on wings, neuron 13860 on antennae, neuron 4040 on thorax, and neuron 16584 on the leg-body junction. Spatial grounding comes from applying the encoder patch-wise, thresholding high activations, forming a mask, and fitting a tight bounding box over the activated region (Pahuja et al., 2 Apr 2026).

Not every latent is retained. Algorithm 1, “Salient Trait Extraction,” uses a species-contrastive criterion. For each image $x_i$ with species label $y_i$, the sparse latent vector is

$z_i = f_\theta(x_i),$

and activated neurons are

$$\mathcal{Z}_i = \{ j \mid z_i[j] > t_{\text{activation}} \},$$

with $$h = W_e (x - b_d) + b_e,\quad g(x) = \text{ReLU}(h),\quad \tilde{x} = W_d\, g(x) + b_d,$$0. Species and genus activation counts are then normalized: $$h = W_e (x - b_d) + b_e,\quad g(x) = \text{ReLU}(h),\quad \tilde{x} = W_d\, g(x) + b_d,$$1 A latent $$h = W_e (x - b_d) + b_e,\quad g(x) = \text{ReLU}(h),\quad \tilde{x} = W_d\, g(x) + b_d,$$2 is retained for species $$h = W_e (x - b_d) + b_e,\quad g(x) = \text{ReLU}(h),\quad \tilde{x} = W_d\, g(x) + b_d,$$3 if

$$h = W_e (x - b_d) + b_e,\quad g(x) = \text{ReLU}(h),\quad \tilde{x} = W_d\, g(x) + b_d,$$4

where $$h = W_e (x - b_d) + b_e,\quad g(x) = \text{ReLU}(h),\quad \tilde{x} = W_d\, g(x) + b_d,$$5 is varied in ablations, including $$h = W_e (x - b_d) + b_e,\quad g(x) = \text{ReLU}(h),\quad \tilde{x} = W_d\, g(x) + b_d,$$6, $$h = W_e (x - b_d) + b_e,\quad g(x) = \text{ReLU}(h),\quad \tilde{x} = W_d\, g(x) + b_d,$$7, and $$h = W_e (x - b_d) + b_e,\quad g(x) = \text{ReLU}(h),\quad \tilde{x} = W_d\, g(x) + b_d,$$8 (Pahuja et al., 2 Apr 2026).

The prompting stage uses a multimodal LLM. The default open model is Qwen2.5-VL-72B, with Qwen2.5-VL-7B and GPT-5-mini used in ablations. The model is given one or several images, typically three, each marked with a red box around the region of interest, and prompted to describe the morphological traits visible within the red boxes and common across all three images. The prompt emphasizes body parts such as wing, leg, antenna, thorax, abdomen, eye, and joint, together with attributes such as color, pattern, shape, size, and texture (Pahuja et al., 2 Apr 2026).

4. Evaluation and ablation results

Evaluation relies primarily on human judgment because there is no large ground-truth insect trait-text benchmark. Three domain-knowledgeable annotators rate sampled descriptions on a 1–5 rubric that scores body-part correctness, attribute correctness, and hallucination. The reported average ratings are normalized to account for rater-specific scale usage (Pahuja et al., 2 Apr 2026).

The main comparisons show that localization is decisive. MLLM-only prompting with three full images yields an average normalized rating of about 3.15, whereas MLLM+SAE with three images yields about 3.91. Using three images is slightly better than using one image: MLLM+SAE with one image scores 3.84, while the three-image version scores 3.91. Model size also matters: within MLLM+SAE, Qwen2.5-VL-7B scores 2.90, Qwen2.5-VL-72B scores 3.58, and GPT-5-mini scores 4.04 (Pahuja et al., 2 Apr 2026).

Ablations on sparse-autoencoder sparsity indicate that less sparse models can be preferable for trait coverage. The configuration with $$h = W_e (x - b_d) + b_e,\quad g(x) = \text{ReLU}(h),\quad \tilde{x} = W_d\, g(x) + b_d,$$9 and $$\mathcal{J}(\phi) = \| x - \tilde{x} \|_2^2 + \alpha\, \mathcal{R}(g(x)).$$0 achieved the best human rating, 3.91. Increasing $$\mathcal{J}(\phi) = \| x - \tilde{x} \|_2^2 + \alpha\, \mathcal{R}(g(x)).$$1 greatly reduces coverage: $$\mathcal{J}(\phi) = \| x - \tilde{x} \|_2^2 + \alpha\, \mathcal{R}(g(x)).$$2 yields 460 images and 7897 traits, $$\mathcal{J}(\phi) = \| x - \tilde{x} \|_2^2 + \alpha\, \mathcal{R}(g(x)).$$3 yields 322 images and 785 traits, and $$\mathcal{J}(\phi) = \| x - \tilde{x} \|_2^2 + \alpha\, \mathcal{R}(g(x)).$$4 yields 20 images and 20 traits (Pahuja et al., 2 Apr 2026).

The baselines also clarify what the pipeline is not doing. Grad-CAM localization is described as diffuse and often blending multiple parts, and MLLM-only prompting tends to produce higher-level full-body descriptions instead of part-focused morphology. The evidence therefore supports the claim that monosemantic spatial grounding is central to Bioscan-Traits rather than incidental.

5. Downstream uses and relation to foundation models

The immediate ecological motivation is to make large-scale morphological analyses possible without expert-only manual annotation. The paper identifies several uses: studying trait distributions across species, relating morphology to climate or habitat, expanding trait databases, and supporting trait-aware identification tools (Pahuja et al., 2 Apr 2026). Because BIOSCAN-5M already contains geography, taxonomy, DNA, and size information, Bioscan-Traits adds an interpretable morphological channel to an already multimodal biodiversity resource (Gharaee et al., 2024).

The dataset is also used as supervision for biodiversity foundation models. Fine-tuning BioCLIP and BioCLIP 2 with Bioscan-Traits improves zero-shot species classification on the “Insects” dataset from Meta-Album, which contains volunteer photos with cluttered backgrounds, occlusion, and distribution shift from lab specimens. Reported accuracies are: BioCLIP baseline 34.8%, BioCLIP with species-level fine-tuning 39.6%, BioCLIP with trait-level fine-tuning 39.9%, BioCLIP 2 baseline 55.3%, and BioCLIP 2 with trait-level fine-tuning 56.23% (Pahuja et al., 2 Apr 2026).

This indicates that Bioscan-Traits is not only a dataset of descriptions but also a mechanism for injecting biologically meaningful supervision into vision-LLMs. A plausible implication is that trait-level captions help models focus on diagnostic morphology instead of background or collection-specific artifacts. That interpretation is consistent with BIOSCAN-5M’s broader emphasis on multimodal representation learning across images, DNA barcodes, and taxonomic text (Gharaee et al., 2024).

6. Limitations, access, and future directions

Several limitations are stated explicitly. The method depends on DINOv2 features encoding morphology well, so backbone bias can suppress certain traits. Monosemanticity is not perfect; some latent units may still correspond to entangled or composite traits. Smaller multimodal LLMs can hallucinate traits, especially when patches include background. Evaluation depends on expert ratings of small samples because no large insect trait-text benchmark exists. Bioscan-Traits also inherits taxonomic, geographic, and imaging biases from BIOSCAN-5M (Pahuja et al., 2 Apr 2026).

The paper also notes methodological caveats around sparse autoencoders. Recent work has criticized SAEs for not always improving downstream performance or for feature hedging; the authors therefore position SAEs here as part detectors rather than as a general-purpose interpretability guarantee. They mitigate these issues through species-contrastive filtering, multi-image prompting, and empirical validation (Pahuja et al., 2 Apr 2026).

Future directions mentioned include extending the pipeline beyond insects to plants, birds, and fungi; using more advanced SAE variants such as BatchTopK and Matryoshka SAEs; integrating explicit trait ontologies or controlled vocabularies; building interactive workflows in which experts refine generated traits; and developing better metrics and benchmarks for trait correctness and ecological utility (Pahuja et al., 2 Apr 2026). Bioscan-Traits and the associated code are openly available via the Hugging Face dataset osunlp/bioscan-traits and the GitHub repository OSU-NLP-Group/sae-trait-annotation, while the underlying BIOSCAN-5M dataset is available through the BIOSCAN-5M project and code repository (Pahuja et al., 2 Apr 2026, Gharaee et al., 2024).

In summary, Bioscan-Traits is best understood as a trait-annotation layer over BIOSCAN-5M: a dataset of 80.8K free-form, region-anchored morphological descriptions for 19.1K insect images, produced by combining DINOv2, sparse autoencoders, species-contrastive selection, and multimodal language prompting. Its significance lies in turning specimen images into trait-bearing records that are interpretable to both ecologists and machine-learning systems, while preserving the scale required for biodiversity analysis (Pahuja et al., 2 Apr 2026).