Model card: v4
Browse files
README.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
-
base_model: intfloat/multilingual-e5-small
|
| 4 |
library_name: coreml
|
| 5 |
pipeline_tag: sentence-similarity
|
| 6 |
language: [en, ru, es]
|
|
@@ -22,6 +22,16 @@ Each zip holds `intfloat.multilingual-e5-small/` with:
|
|
| 22 |
|
| 23 |
Prefix texts with `query: ` or `passage: ` as the original model card says.
|
| 24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
### v3: v2's data plus synthetic personal documents
|
| 26 |
|
| 27 |
`e5-small-coreml-v3.zip` (112,125,116 bytes, SHA-256
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
+
base_model: [intfloat/multilingual-e5-small, intfloat/multilingual-e5-base]
|
| 4 |
library_name: coreml
|
| 5 |
pipeline_tag: sentence-similarity
|
| 6 |
language: [en, ru, es]
|
|
|
|
| 22 |
|
| 23 |
Prefix texts with `query: ` or `passage: ` as the original model card says.
|
| 24 |
|
| 25 |
+
### v4: multilingual-e5-base, same fine-tune data as v3
|
| 26 |
+
|
| 27 |
+
`e5-base-coreml-v4.zip` (260,538,460 bytes, SHA-256
|
| 28 |
+
`d20e55a82a06ce668543ac95cab56067402391eab5ed63bd7febc3219361ee63`). Base model
|
| 29 |
+
[intfloat/multilingual-e5-base](https://huggingface.co/intfloat/multilingual-e5-base) (MIT, 278M parameters,
|
| 30 |
+
768-dimensional output; the folder name stays `intfloat.multilingual-e5-small` for the app). Same vocabulary.
|
| 31 |
+
v3's training pairs (114,268), batch 32 x 3500 steps, lr 1e-5. Core ML int8 vs PyTorch: cosine >= 0.998.
|
| 32 |
+
Private held-out set of 44 file-search queries over real personal documents, whole files: right file first
|
| 33 |
+
20 -> 26 (v3 -> v4), MRR 0.637 -> 0.720. Belebele ru/es/en cross-language hit@1 0.786 -> 0.826.
|
| 34 |
+
|
| 35 |
### v3: v2's data plus synthetic personal documents
|
| 36 |
|
| 37 |
`e5-small-coreml-v3.zip` (112,125,116 bytes, SHA-256
|