Downloads · 30 days
13
18% of all-time downloads
tuanda2912/hark-multilingual-e5-small-coreml
hark-multilingual-e5-small-coreml is a sentence similarity model from tuanda2912. Use it when you need a score for how close two texts are. It is set up for coreml. The card lists the license as mit.
A CoreML (mlprogram, int8-weight-quantized) conversion of intfloat/multilingual-e5-small (384-dim, multilingual), packaged for on-device vault search in Hark — a local-first, macOS-only meeting transcription app. Runs…
Downloads · 30 days
13
18% of all-time downloads
All-time downloads
71
Public
Repo size
282 MB
Likes
0
Public
Click a slice to open those files.
.bin118 MB · 84%
From the Hugging Face model README
A CoreML (mlprogram, int8-weight-quantized) conversion of
intfloat/multilingual-e5-small
(384-dim, multilingual), packaged for on-device vault search in
Hark — a local-first, macOS-only meeting
transcription app. Runs on the Apple Neural Engine; nothing is sent off the
machine (Hark embeds the whole vault locally).
This repo exists so Hark can download a ready-to-run CoreML artifact instead of shipping it in the app bundle. It is a faithful conversion — see Provenance and Validation — not a new model.
| File | What |
|---|---|
MultilingualE5Small.mlpackage/ | the CoreML model (int8 weights, ~113 MB) |
tokenizer.json, tokenizer_config.json, special_tokens_map.json | XLM-RoBERTa tokenizer (SentencePiece Unigram) |
sentencepiece.bpe.model | the SentencePiece model |
Hark's loader snapshots this repo at a pinned revision into its app-support
models dir, compiles the .mlpackage to the ANE, and runs fully offline
thereafter.
input_ids (int32 [1, L]), attention_mask (int32 [1, L]), flexible L ∈ 1..512last_hidden_state (float32 [1, L, 384])"query: " / "passage: "). Reproduce those if you reuse
this model directly.intfloat/multilingual-e5-small at source revision
614241f622f53c4eeff9890bdc4f31cfecc418b3 via
engine/scripts/convert-embedder-coreml.py
(coremltools 9, convert_to="mlprogram", minimum_deployment_target=macOS14).engine/scripts/quantize-embedder-int8.py
(coremltools.optimize.coreml.linear_quantize_weights).MIT, inherited from intfloat/multilingual-e5-small. This is a format conversion