Update README.md

README.md

@@ -5,16 +5,15 @@ language:
 datasets:
 - MIRACL
 tags:
-- miniMiracle
 - miniDense
 - passage-retrieval
 - knowledge-distillation
 - middle-training
 - sentence-transformers
 pretty_name: >-
-  miniMiracle is a family of High-quality, Light Weight and Easy deploy
+  miniDense is a family of High-quality, Light Weight and Easy deploy
   multilingual embedders / retrievers, primarily focussed on Indo-Aryan and
-  Indo-Dravidin Languages.
+  Indo-Dravidian Languages.
 library_name: transformers
 pipeline_tag: sentence-similarity
 ---
@@ -64,10 +63,10 @@ pipeline_tag: sentence-similarity
 
 # Request, Terms, Disclaimers
 
-https://github.com/sponsors/PrithivirajDamodaran 
 
 <center>
   <img src="./ar_terms.png" width=250%/>
+  <b><p>[https://github.com/sponsors/PrithivirajDamodaran](https://github.com/sponsors/PrithivirajDamodaran)</p><b>
 </center>
 
 
@@ -178,20 +177,31 @@ The below numbers are with mDPR model, but miniDense_arabic_v1 should give a eve
 
 *Note: MIRACL paper shows a different (higher) value for BM25 Arabic, So we are taking that value from BGE-M3 paper, rest all are form the MIRACL paper.*
 
-#### cMTEB numbers:
-CMTEB is a general purpose embedding evaluation benchmark covering wide range of tasks, but like BGE-M3, miniMiracle models are predominantly tuned for retireval tasks aimed at search & IR based usecases.
-We ran the retrieval slice of the cMTEB.
+#### MTEB numbers:
+MTEB is a general purpose embedding evaluation benchmark covering wide range of tasks, but miniDense models (like BGE-M3) are predominantly tuned for retireval tasks aimed at search & IR based usecases.
+So it makes sense to evaluate our models in retrieval slice of the MTEB benchmark.
 
-We compared the performance few top general purpose embedding models on the C-MTEB benchmark. please refer to the C-MTEB leaderboard. Almost all models below are bert-Arabic based so they have no notion of any other languages.
+##### Long Document Retrieval 
 
 <center>
-  <img src="./ar_metrics_3.png" width=150%/>
+<img src="./ar_metrics_4.png" width=100%/>
+  <b><p>Table 3: Detailed Arabic retrieval performance on the MultiLongDoc dev set (measured by nDCG@10)</p></b>
+</center>
+
+
+##### X-lingual Retrieval 
+
+Almost all models below are monolingual arabic models based so they have no notion of any other languages.
+
+<center>
+<img src="./ar_metrics_5.png" width=100%/>
+  <b><p>Table 4: Detailed Arabic retrieval performance on the 3 X-lingual test set (measured by nDCG@10)</p></b>
 </center>
 
 <br/>
 
 # Roadmap
-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
+We will add miniDense 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
 
 - Spanish
 - Tamil
@@ -203,7 +213,7 @@ We will add miniMiracle series of models for all popular languages as we see fit
 
 We welcome anyone to reproduce our results. Here are some tips and observations:
 
-- Use CLS Pooling and Inner Product.
+- Use CLS Pooling (not mean) and Inner Product (not cosine).
 - There *may be* minor differences in the numbers when reproducing, for instance BGE-M3 reports a nDCG@10 of 59.3 for MIRACL hindi and we Observed only 58.9.
 
 Here are our numbers for the full hindi run on BGE-M3