hillbyte/labse-mizo-en-embed
This model is a fine-tuned version of sentence-transformers/LaBSE on the Mizo-English parallel corpus 170k.
It maps both English and Mizo (Lushai) sentences into a shared 768-dimensional dense vector space, enabling state-of-the-art cross-lingual semantic search, information retrieval (CLIR/CLRI), bitext alignment/mining, and sentence similarity.
📊 Cross-Lingual Information Retrieval (CLRI) Benchmark
Evaluated on 3,000 holdout test query-document pairs
🔹 English Query -> Mizo Corpus Search
| Metric | Baseline (Pretrained LaBSE) | Fine-Tuned Model | Absolute Improvement |
|---|---|---|---|
| nDCG@10 | 54.02% | 98.24% | +44.23% |
| MRR@10 | 51.46% | 97.93% | +46.47% |
| Recall@1 (Top-1) | 46.70% | 97.07% | +50.37% |
| Recall@5 | 57.63% | 98.93% | +41.30% |
| Recall@10 | 62.17% | 99.17% | +37.00% |
| MAP@100 | 51.98% | 97.96% | +45.98% |
🔹 Mizo Query -> English Corpus Search
| Metric | Baseline (Pretrained LaBSE) | Fine-Tuned Model | Absolute Improvement |
|---|---|---|---|
| nDCG@10 | 52.20% | 97.97% | +45.77% |
| MRR@10 | 48.97% | 97.61% | +48.64% |
| Recall@1 (Top-1) | 43.07% | 96.73% | +53.67% |
| Recall@5 | 56.57% | 98.67% | +42.10% |
| Recall@10 | 62.53% | 99.07% | +36.53% |
| MAP@100 | 49.61% | 97.64% | +48.03% |
🚀 Quick Start & Usage
from sentence_transformers import SentenceTransformer, util
# Load the fine-tuned model
model = SentenceTransformer("hillbyte/labse-mizo-en-embed")
# Example bilingual queries & targets
en_sentences = [
"Jehovah expects us to render sacred service to him in a clean manner.",
"The weather in Aizawl is pleasant today."
]
mz_sentences = [
"Jehova'n thianghlim taka a rawngbawl turin min beisei a.",
"Vawiin chu Aizawl khua a nuam hle mai."
]
# Compute embeddings
en_embeddings = model.encode(en_sentences, convert_to_tensor=True)
mz_embeddings = model.encode(mz_sentences, convert_to_tensor=True)
# Compute cosine similarity matrix
cosine_scores = util.cos_sim(en_embeddings, mz_embeddings)
print("Cosine Similarity Matrix:")
print(cosine_scores)
⚙️ Training Details
- Base Model:
sentence-transformers/LaBSE(12-layer Multilingual BERT, 768 hidden size) - Dataset: ~170,817 parallel training pairs
- Loss Function:
MultipleNegativesRankingLosswith in-batch negatives - Batch Size: 64 (producing 63 in-batch negatives per sample)
- Learning Rate:
2e-5with linear warmup & decay - Epochs: 3 (8,010 total steps)