You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Export support is fragmented across openmed/mlx (bert/deberta/privacy-filter/gliner families, INT8/INT4), openmed/coreml (float16/32, token-classification only), and the new ONNX/WebGPU/Transformers.js/GGUF paths, with real limitations (GGUF embeddings only; CoreML token-classification only; MLX family allowlist). There is no single doc telling contributors which (architecture x format x tier) combinations are supported, so users guess and file issues. This is a self-contained docs task with a coherence test, ideal for a first contribution.
Scope
Author docs/export-matrix.md as a table of architecture-family rows (BERT/DistilBERT/Electra/RoBERTa/XLM-R/DeBERTa-v2/ModernBERT/Longformer/privacy-filter/GLiNER) by export format columns (MLX-fp, MLX-8bit, MLX-4bit, CoreML-fp16, CoreML-int8, ONNX, Transformers.js, GGUF), marking supported / partial / unsupported with the reason.
Capture the known limitations verbatim: GGUF is embedding-only, CoreML is token-classification-only, the MLX family allowlist, INT4 only-if-recall-holds.
Add a small test that parses export-matrix.md and asserts every family in the MLX _SUPPORTED_TOKEN_CLASSIFICATION_MODEL_TYPES allowlist has a row (keeps the doc honest as backends are added).
Link the matrix from the existing export docs.
Acceptance criteria
docs/export-matrix.md exists with an architecture-by-format support table and per-cell reasons for partial/unsupported.
A test asserts every MLX-supported family appears as a row in the matrix.
The matrix captures the GGUF-embedding-only and CoreML-token-classification-only limitations.
test suite green: .venv/bin/python -m pytest tests/ -q
Out of scope
Implementing any new export backend.
Auto-generating the matrix from code (a static doc + coherence test is sufficient).
Files
docs/export-matrix.md
tests/unit/test_export_matrix.py
Task: OM-146 · Milestone: Backlog · Priority: P3 · Size: S
Depends on: — · Blocks: —
Roadmap: section 2.2c; Appendix C (export file index)
Spec: PLANS/V2/EXECUTION/tasks/OM-146.md
Summary
Export support is fragmented across openmed/mlx (bert/deberta/privacy-filter/gliner families, INT8/INT4), openmed/coreml (float16/32, token-classification only), and the new ONNX/WebGPU/Transformers.js/GGUF paths, with real limitations (GGUF embeddings only; CoreML token-classification only; MLX family allowlist). There is no single doc telling contributors which (architecture x format x tier) combinations are supported, so users guess and file issues. This is a self-contained docs task with a coherence test, ideal for a first contribution.
Scope
Acceptance criteria
Out of scope
Files
Task: OM-146 · Milestone: Backlog · Priority: P3 · Size: S
Depends on: — · Blocks: —
Roadmap: section 2.2c; Appendix C (export file index)
Spec: PLANS/V2/EXECUTION/tasks/OM-146.md