Model Card for SSEL

Checkpoints and extracted DNA embeddings for Hierarchical Embedding of DNA Barcode Taxonomy with Hyperbolic Entailment Cones. This repository contains the nine 768D experiments presented in the iBOL poster; eight also appear in the UniReps submission.

Model Details

Model Description

Structured Stacked Entailment Loss (SSEL) extends Stacked Entailment Loss (SEL) with minimum-distance, full-cone nesting, sibling-cone separation and descendant-based cone allocation constraints. DNA barcodes and taxonomic labels are embedded in a shared hyperbolic space.

The DNA branch starts from BarcodeBERT, followed by a 768 → 1536 → 768 MLP without a residual connection. Taxonomic labels use learned embedding tables. These runs use DNA and four taxonomic ranks, without an image branch.

Model Sources

• Project page
• Code preparation repository (available soon)
• Paper

Model Checkpoints

Method	Checkpoint (last.pth)	Test embeddings	Saved configuration	UniReps
EL	el	DNA	Config	Yes
SEL	sel	DNA	Config	Yes
Positive EL + CL	positive-el-cl	DNA	Config	No
SSEL	ssel	DNA	Config	Yes
No direct placement	no-direct-placement	DNA	Config	Yes
Frozen labels	frozen-labels	DNA	Config	Yes
SEL + nesting	sel-nesting	DNA	Config	Yes
SEL + separation	sel-separation	DNA	Config	Yes
SEL + allocation	sel-allocation	DNA	Config	Yes

Each selected checkpoint is the exact last.pth used for its reported result. All nine runs use 768 dimensions, 15 epochs and seed 42. EL and SEL omit minimum distance; the other listed variants include it. Positive EL + CL is our DNA–taxonomy adaptation, not an exact reproduction of image–text MERU.

Files and Usage

All versioned artifacts are under Ver_1_0/. The layout follows CLIBD's ckpt/ and extracted_embedding/ structure, with the version directory placed at the outside:

Ver_1_0/
  ckpt/
    bioscan_clip/bioscan_5m/<model_output_name>/
      last.pth
      config.yaml
    BarcodeBERT/old_checkpoints/trained_with_canada_1_5M/model_41.pth
  extracted_embedding/
    bioscan_5m/<model_output_name>/
      extracted_feature_from_test_split.hdf5
      labels_test.json
      processed_id_test.json
      dna_to_label.csv
      species_detection.csv
      evaluation_complete.json
    training_seen_taxonomy_768d_weighted_simplex_beta025_margin05_gap05_placement/
  data/BIOSCAN_5M/global_label_to_id_mapping.json
  metadata/experiments.json
  SHA256SUMS.txt

The checkpoints are native PyTorch state dictionaries and use the SSEL codebase's loader. They are not Transformers AutoModel packages. Configuration copies retain the scientific settings; local absolute roots become ${oc.env:SSEL_ARTIFACT_ROOT,.}, and W&B is disabled by default. Set SSEL_ARTIFACT_ROOT to the downloaded Ver_1_0 directory. The processed BIOSCAN-5M HDF5 is downloaded separately into data/BIOSCAN_5M/ within that root. The current loader additionally uses prajjwal1/bert-small from Hugging Face or its local cache.

Embedding Format

Each HDF5 contains test DNA embeddings only, in the saved checkpoint's hyperbolic space:

HDF5 dataset	Shape	Type
seen/encoded_dna_feature	39,373 × 768	float32
unseen/encoded_dna_feature	7,887 × 768	float32

The companion JSON files provide labels and process IDs in exactly the same row order. These are spatial coordinates of Lorentz-model hyperbolic embeddings; they are not unit-normalized Euclidean vectors. Recover the time coordinate using the checkpoint curvature when needed.

Final trained taxonomic label tables, learned scales and curvature are included inside each checkpoint. The shared placement artifact supplies initial coordinates, candidate labels, global IDs and parent relationships; initial coordinates are not final trained label embeddings. Separate final-label vector files, train/validation embeddings and specimen reference-key caches are not included in this version.

Classification uses the nearest label by exterior angle at each rank. Unseen-species detection uses cone checks against the seen-species label candidates. These are separate measurements; the accompanying metric CSVs and evaluation records preserve that distinction.

Training Data

The experiments use BIOSCAN-5M. The processed HDF5 has 39,373 seen-species and 7,887 unseen-species test queries. Label candidates are training-derived: 55 orders, 929 families, 6,819 genera and 11,846 species. The raw dataset is not stored in this model repository.

Acknowledgement

This work was supported by the Government of Canada's New Frontiers in Research Fund (NFRF) [NFRFT-2020-00073] and the Canada CIFAR AI Chairs Program. We acknowledge BIOSCAN-5M, BarcodeBERT and the CLIBD codebase. This model card and artifact layout follow CLIBD.