Table 1: Hindi retrieval performance on the MIRACL dev set (measured by nDCG@10)

Architecture:

Model: BERT.
Tokenizer: XLM-Roberta's Tokenizer.

License and Terms:
Detailed comparison & Our Contribution:
ONNX & GGUF Status:
Usage:
- With Sentence Transformers:
- With Huggingface Transformers:
FAQs
Roadmap
Notes on Reproducing:
Reference:
Note on model bias

License and Terms:

Detailed comparison & Our Contribution:

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. We will add more details here.

Table 2: Detailed Hindi retrieval performance on the MIRACL dev set (measured by nDCG@10)

Full set of evaluation numbers for our model

{'NDCG@1': 0.42571, 'NDCG@3': 0.42062, 'NDCG@5': 0.44842, 'NDCG@10': 0.5039, 'NDCG@100': 0.56175, 'NDCG@1000': 0.57772}
{'MAP@1': 0.22683, 'MAP@3': 0.33514, 'MAP@5': 0.37345, 'MAP@10': 0.40861, 'MAP@100': 0.42833, 'MAP@1000': 0.42916}
{'Recall@10': 0.63964, 'Recall@50': 0.80537, 'Recall@100': 0.87136, 'Recall@200': 0.9211, 'Recall@500': 0.96851, 'Recall@1000': 0.97987}
{'P@1': 0.42571, 'P@3': 0.27429, 'P@5': 0.212, 'P@10': 0.13943, 'P@100': 0.01911, 'P@1000': 0.00211}
{'MRR@10': 0.53057, 'MRR@100': 0.53736, 'MRR@1000': 0.5377}

ONNX & GGUF Status:

Variant	Status
FP16 ONNX	✅
GGUF	WIP

Usage:

With Sentence Transformers:

from sentence_transformers import SentenceTransformer
import scipy.spatial


model = SentenceTransformer('prithivida/miniMiracle_hi_v1')

corpus = [
    'एक आदमी खाना खा रहा है।',
    'लोग ब्रेड का एक टुकड़ा खा रहे हैं।',
    'लड़की एक बच्चे को उठाए हुए है।',
    'एक आदमी घोड़े पर सवार है।',
    'एक महिला वायलिन बजा रही है।',
    'दो आदमी जंगल में गाड़ी धकेल रहे हैं।',
    'एक आदमी एक सफेद घोड़े पर एक बंद मैदान में सवारी कर रहा है।',
    'एक बंदर ड्रम बजा रहा है।',
    'एक चीता अपने शिकार के पीछे दौड़ रहा है।',
    'एक बड़ा डिनर है।'
]

corpus_embeddings = model.encode(corpus)

queries = [
    'एक आदमी पास्ता खा रहा है।',
    'एक गोरिल्ला सूट पहने व्यक्ति ड्रम बजा रहा है।'
]

query_embeddings = model.encode(queries)

# Find the closest 3 sentences of the corpus for each query sentence based on cosine similarity
closest_n = 3
for query, query_embedding in zip(queries, query_embeddings):
    distances = scipy.spatial.distance.cdist([query_embedding], corpus_embeddings, "cosine")[0]

    results = zip(range(len(distances)), distances)
    results = sorted(results, key=lambda x: x[1])

    print("\n======================\n")
    print("Query:", query)
    print("\nTop 3 most similar sentences in corpus:\n")

    for idx, distance in results[0:closest_n]:
        print(corpus[idx].strip(), "(Score: %.4f)" % (1-distance))

# Optional: How to quantize the embeddings
# binary_embeddings = quantize_embeddings(embeddings, precision="ubinary")

With Huggingface Transformers:

T.B.A

FAQs:

How can I reduce overall inference cost ?

You can host these models without heavy torch dependency using the ONNX flavours of these models via FlashEmbed library.

How do I reduce vector storage cost ?

Use Binary and Scalar Quantisation

How do I offer hybrid search to improve accuracy ?

MIRACL paper shows simply combining BM25 is a good starting point for a Hybrid option: The below numbers are with mDPR model, but miniMiracle_hi_v1 should give a even better hybrid performance.

Language	ISO	nDCG@10 BM25	nDCG@10 mDPR	nDCG@10 Hybrid
Hindi	hi	0.458	0.383	0.616

Why not run MTEB?

MTEB is a general purpose embedding evaluation bechmark covering wide range of tasks available currently only for English, Chinese, French and few other languages but not Indic languages. Besides like BGE-M3, miniMiracle models are predominantly tuned for retireval tasks aimed at search & IR based usecases. At the moment MIRACL is the gold standard for a subset of Indic languages.

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

Spanish
Tamil
Arabic
German
English ?

Notes on reproducing:

We welcome everyone to reproduce our results. Here are some tips and observations:

Use CLS Pooling and Inner Product.
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

{'NDCG@1': 0.49714, 'NDCG@3': 0.5115, 'NDCG@5': 0.53908, 'NDCG@10': 0.58936, 'NDCG@100': 0.6457, 'NDCG@1000': 0.65336}
{'MAP@1': 0.28845, 'MAP@3': 0.42424, 'MAP@5': 0.46455, 'MAP@10': 0.49955, 'MAP@100': 0.51886, 'MAP@1000': 0.51933}
{'Recall@10': 0.73032, 'Recall@50': 0.8987, 'Recall@100': 0.93974, 'Recall@200': 0.95763, 'Recall@500': 0.97813, 'Recall@1000': 0.9902}
{'P@1': 0.49714, 'P@3': 0.33048, 'P@5': 0.24629, 'P@10': 0.15543, 'P@100': 0.0202, 'P@1000': 0.00212}
{'MRR@10': 0.60893, 'MRR@100': 0.615, 'MRR@1000': 0.6151}

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.

Reference:

Note on model bias:

Like any model this model might carry inherent biases from the base models and the datasets it was pretrained and finetuned on. Please use responsibly.

prithivida
/

miniMiracle_hi_v1