Cliphoard Core ML model conversions
Existing Core ML conversions used by Cliphoard, mirrored byte-for-byte from its public models-v1 release. These are conversions of the attributed upstream models, not new base models trained by Axiotic. No executable app or native library is included. Download only the selected model; verify the complete ZIP SHA256 in ARTIFACTS.json before safe extraction and local Core ML compilation.
MiniLM
coreml/all-MiniLM-L6-v2.zip contains the all-MiniLM-L6-v2.mlpackage and tokenizer. Upstream: sentence-transformers/all-MiniLM-L6-v2, Apache2.0. The existing conversion targets macOS13 with INT32 input_ids and attention_mask, dynamic sequence length2–511 and an L2-normalized384-dimensional embedding output. See Cliphoard's Tools/convert_minilm.py; this mirror does not introduce a new conversion or new quality benchmark.
OpenVision Tiny
coreml/openvision-tiny-p8.zip contains separate image/text .mlpackage towers, WordPiece tokenizer and the existing conversion manifest. Upstream: UCSC-VLAA/openvision-vit-tiny-patch8-224, Apache2.0. Existing manifest:192-dimensional joint embedding,224px image,80-token context, L2-normalization and recorded image preprocessing/text-mean adjustment. Preserve that manifest with both towers.
Provenance and limitations
The fixed original archives are at Cliphoard models-v1. ARTIFACTS.json records original archive IDs, hashes, sizes and member inventories. Current upstream repository revisions used to inspect declared licences are not asserted to be the original conversion's weight revision. Historical parity/runtime evidence belongs to Cliphoard; this mirror does not claim new inference validation, exhaustive privacy review or compatibility beyond its original conversion contract.
See LICENSE and NOTICE.md. Model data is supplied under Apache2.0; app runtime/library licensing remains separate.