prithivida
commited on
Commit
•
9a079cf
1
Parent(s):
0e7ac0a
Update README.md
Browse files
README.md
CHANGED
@@ -37,27 +37,31 @@ pipeline_tag: sentence-similarity
|
|
37 |
</center>
|
38 |
|
39 |
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
- [With Sentence Transformers:](#with-sentence-transformers)
|
45 |
- [With Huggingface Transformers:](#with-huggingface-transformers)
|
|
|
|
|
46 |
- [How do I optimise vector index cost?](#how-do-i-optimise-vector-index-cost)
|
47 |
- [How do I offer hybrid search to address Vocabulary Mismatch Problem?](#how-do-i-offer)
|
|
|
|
|
48 |
- [Notes on Reproducing:](#notes-on-reproducing)
|
49 |
- [Reference:](#reference)
|
50 |
- [Note on model bias](#note-on-model-bias)
|
51 |
|
52 |
|
53 |
-
|
54 |
|
55 |
<center>
|
56 |
<img src="./terms.png" width=200%/>
|
57 |
</center>
|
58 |
|
59 |
|
60 |
-
|
61 |
|
62 |
English language famously have **all-minilm** series models which were great for quick experimentations and for certain production workloads. The Idea is to have same for the other popular langauges, starting with Indo-Aryan and Indo-Dravidian languages. Our innovation is in bringing high quality models which easy to serve and embeddings are cheaper to store without ANY pretraining or expensive finetuning. For instance, **all-minilm** are finetuned on 1-Billion pairs. We offer a very lean model but with a huge vocabulary - around 250K.
|
63 |
We will add more details here.
|
@@ -81,7 +85,7 @@ Full set of evaluation numbers for our model
|
|
81 |
|
82 |
<br/>
|
83 |
|
84 |
-
|
85 |
|
86 |
#### With Sentence Transformers:
|
87 |
|
@@ -138,10 +142,16 @@ for query, query_embedding in zip(queries, query_embeddings):
|
|
138 |
#### With Huggingface Transformers:
|
139 |
- T.B.A
|
140 |
|
|
|
|
|
|
|
|
|
|
|
|
|
141 |
#### How do I optimise vector index cost ?
|
142 |
[Use Binary and Scalar Quantisation](https://huggingface.co/blog/embedding-quantization)
|
143 |
|
144 |
-
|
145 |
MIRACL paper shows simply combining BM25 is a good starting point for a Hybrid option:
|
146 |
The below numbers are with mDPR model, but miniMiracle_zh_v1 should give a even better hybrid performance.
|
147 |
|
@@ -151,6 +161,21 @@ The below numbers are with mDPR model, but miniMiracle_zh_v1 should give a even
|
|
151 |
|
152 |
*Note: MIRACL paper shows a different (higher) value for BM25 Chinese, So we are taking that value from BGE-M3 paper, rest all are form the MIRACL paper.*
|
153 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
154 |
# Notes on reproducing:
|
155 |
|
156 |
We welcome anyone to reproduce our results. Here are some tips and observations:
|
@@ -168,7 +193,7 @@ Here are our numbers for the full hindi run on BGE-M3
|
|
168 |
{'MRR@10': 0.60893, 'MRR@100': 0.615, 'MRR@1000': 0.6151}
|
169 |
```
|
170 |
|
171 |
-
Fair warning BGE-M3 is $ expensive to evaluate, probably that's why it's not part of any of the MTEB benchmarks.
|
172 |
|
173 |
|
174 |
# Reference:
|
|
|
37 |
</center>
|
38 |
|
39 |
|
40 |
+
- [License and Terms:](#license-and-terms)
|
41 |
+
- [Detailed comparison & Our Contribution:](#detailed-comparison--our-contribution)
|
42 |
+
- [ONNX & GGUF Variants:](#detailed-comparison--our-contribution)
|
43 |
+
- [Usage:](#usage)
|
44 |
- [With Sentence Transformers:](#with-sentence-transformers)
|
45 |
- [With Huggingface Transformers:](#with-huggingface-transformers)
|
46 |
+
- [FAQs](#faqs)
|
47 |
+
- [How can we run these models with out heavy torch dependency?](#how-can-we-run-these-models-with-out-heavy-torch-dependency)
|
48 |
- [How do I optimise vector index cost?](#how-do-i-optimise-vector-index-cost)
|
49 |
- [How do I offer hybrid search to address Vocabulary Mismatch Problem?](#how-do-i-offer)
|
50 |
+
- [Why not run cMTEB?](#why-not-run-cmteb)
|
51 |
+
- [Roadmap](#roadmap)
|
52 |
- [Notes on Reproducing:](#notes-on-reproducing)
|
53 |
- [Reference:](#reference)
|
54 |
- [Note on model bias](#note-on-model-bias)
|
55 |
|
56 |
|
57 |
+
# License and Terms:
|
58 |
|
59 |
<center>
|
60 |
<img src="./terms.png" width=200%/>
|
61 |
</center>
|
62 |
|
63 |
|
64 |
+
# Detailed comparison & Our Contribution:
|
65 |
|
66 |
English language famously have **all-minilm** series models which were great for quick experimentations and for certain production workloads. The Idea is to have same for the other popular langauges, starting with Indo-Aryan and Indo-Dravidian languages. Our innovation is in bringing high quality models which easy to serve and embeddings are cheaper to store without ANY pretraining or expensive finetuning. For instance, **all-minilm** are finetuned on 1-Billion pairs. We offer a very lean model but with a huge vocabulary - around 250K.
|
67 |
We will add more details here.
|
|
|
85 |
|
86 |
<br/>
|
87 |
|
88 |
+
# Usage:
|
89 |
|
90 |
#### With Sentence Transformers:
|
91 |
|
|
|
142 |
#### With Huggingface Transformers:
|
143 |
- T.B.A
|
144 |
|
145 |
+
# FAQs:
|
146 |
+
|
147 |
+
#### How can we run these models with out heavy torch dependency?
|
148 |
+
|
149 |
+
- You can use ONNX flavours of these models via [FlashRetrieve](https://github.com/PrithivirajDamodaran/FlashRetrieve) library.
|
150 |
+
|
151 |
#### How do I optimise vector index cost ?
|
152 |
[Use Binary and Scalar Quantisation](https://huggingface.co/blog/embedding-quantization)
|
153 |
|
154 |
+
#### How do I offer hybrid search to address Vocabulary Mismatch Problem?
|
155 |
MIRACL paper shows simply combining BM25 is a good starting point for a Hybrid option:
|
156 |
The below numbers are with mDPR model, but miniMiracle_zh_v1 should give a even better hybrid performance.
|
157 |
|
|
|
161 |
|
162 |
*Note: MIRACL paper shows a different (higher) value for BM25 Chinese, So we are taking that value from BGE-M3 paper, rest all are form the MIRACL paper.*
|
163 |
|
164 |
+
#### Why not run cMTEB?
|
165 |
+
CMTEB is a general purpose embedding evaluation bechmark covering wide range of tasks, but like BGE-M3, miniMiracle models are predominantly tuned for retireval tasks aimed at search & IR based usecases.
|
166 |
+
But we would run the retrieval slice of the cMTEB and add the scores here.
|
167 |
+
|
168 |
+
|
169 |
+
# Roadmap
|
170 |
+
We will add miniMiracle series of models for all popular languages as we see fit or based on community requests in phases. Some of the languages we have in our list are
|
171 |
+
|
172 |
+
- Spanish
|
173 |
+
- Tamil
|
174 |
+
- Arabic
|
175 |
+
- German
|
176 |
+
- English ?
|
177 |
+
|
178 |
+
|
179 |
# Notes on reproducing:
|
180 |
|
181 |
We welcome anyone to reproduce our results. Here are some tips and observations:
|
|
|
193 |
{'MRR@10': 0.60893, 'MRR@100': 0.615, 'MRR@1000': 0.6151}
|
194 |
```
|
195 |
|
196 |
+
Fair warning BGE-M3 is $ expensive to evaluate, probably that's why it's not part of any of the retrieval slice of MTEB benchmarks.
|
197 |
|
198 |
|
199 |
# Reference:
|