Semantic-Swingers Task A, B, and C, at LLMs4OL-2026

[matias-vizcaino]

Team: Semantic-Swingers

Tasks included: Task A — Flagship (Text2Onto), Task B — Reuse (Term Typing), Task C — Taxonomy Discovery

Submission type: Native OntoLearner integration (Option 3).

Pull Request (native integration): #338

GitHub Repo (public fork / branch): https://github.com/matias-vizcaino/OntoLearner/tree/feat/semanticswingers-llms4ol2026

Fine-tuned weights (Hugging Face Hub, public):

• https://huggingface.co/datagero/qwen3.5-9b-ontology-extraction-raft — Task A, retrieval-aware FT (k=10), CUDA/peft
• https://huggingface.co/datagero/qwen3.5-9b-ontology-extraction-baseft — Task A, base FT (k=0), CUDA/peft
• https://huggingface.co/datagero/qwen3.5-9b-ontology-extraction-baseft-mlx — Task A, base FT, Apple-Silicon/MLX
• https://huggingface.co/datagero/taxonomy-structural-matrix-1024-mxbai — Task C, 1024-D Bilinear Adjacency Layer adapter (mxbai-embed-large-v1)

What we integrated: AutoLearner subclasses for all three tasks (_taxonomy_discovery / _term_typing / _text2onto), each validated end-to-end through LearnerPipeline and native metrics. Backends support peft/mlx/ollama/openai.

For Task C specifically, we integrated a hybrid pipeline featuring a 1024-D Structural Matrix trained via contrastive margin-ranking to eliminate sibling-edges, coupled with NetworkX DAG cycle-breaking, transitive reduction, and a scale-aware automatic fallback for massive ontologies. This runs as a scalable alternative alongside our primary LLM taxonomy pipeline.

Quality gates: ruff (v0.8.2) clean + pytest green (63 passing). Read-the-docs page and runnable examples included. See PR #338 for details, including a "Future work / discussion" section proposing a document-grounded taxonomy-discovery variant and a triple-preserving text2onto + graph_similarity metric. We are available during the review process.

Supplementary: an end-to-end reproduction notebook (all three tasks) and a methodological diagram are available on a companion branch of our fork — https://github.com/matias-vizcaino/OntoLearner/tree/companion/semanticswingers-llms4ol2026 (see notebooks/README.md for context and how to run, and METHODOLOGY.md for the diagram). It is reference material only, not part of PR #338.

[matias-vizcaino]

Supplementary: an end-to-end reproduction notebook (all three tasks) is available on a companion branch of our fork — https://github.com/matias-vizcaino/OntoLearner/tree/companion/semanticswingers-llms4ol2026 (see notebooks/README.md for context and how to run). It is reference material only, not part of PR #338.

[HamedBabaei]

Dear @matias-vizcaino , thanks for your submission and for the effort on the integration. I will mark your submission as completed. We will use this issue to track your PR, and aslo we will also get back to your PR code review.

[HamedBabaei]

System Integration Submission Checklist

Team Name: Semantic-Swingers
Challenge Variant: LLMs4OL-2026 Flagship, Reuse, and Taxonomy
Submission Type: Native OntoLearner Integration

---

General Requirements

• Team name and challenge variant are included in the issue (or PR) title.
• The submission is complete and executable.
• Unit tests are included for the core system components.
• A methodological diagram is included.
• A link to the project GitHub repository is included.

Native OntoLearner Integration

• Forked the OntoLearner repository.
• System is integrated following the OntoLearner integration guide.
• Commit guidelines have been followed.
• Read-the-Docs documentation has been added.
• Unit tests have been added.
• An example run has been added to the examples/ directory.
• Pull Request title includes the team name and challenge variant.
• Available to address review comments during the PR review process.

---

Pull Request for review: #338