File size: 3,638 Bytes
daf11e2 259d7d9 b458fe2 259d7d9 daf11e2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 |
---
license: mit
task_categories:
- text-generation
- translation
size_categories:
- 100K<n<1M
---
## 🧠 Dataset Card: Embedder — Multilingual Triplet Embedding Dataset
### 📌 Overview
**Embedder** is a multilingual triplet dataset designed for training and evaluating sentence embedding models using contrastive or triplet loss. It contains 1m examples across 11 Indic languages and English + 100extra langs, derived from the Samanantar parallel corpus and opus 100. Each example is structured as a triplet: `(anchor, positive, negative)`.
This dataset is ideal for building cross-lingual retrieval systems, semantic search engines, and multilingual embedding models.
---
### 🏗️ Construction Details
The dataset was built using the following pipeline:
- **Source**: [AI4Bharat Samanantar](https://huggingface.co/datasets/ai4bharat/samanantar) — a high-quality parallel corpus for 11 Indic languages ↔ English.
- **Step 1: Sampling**
Randomly sampled bilingual sentence pairs from Samanantar, ensuring diverse language coverage and semantic alignment.
- **Step 2: Triplet Formation**
- `anchor`: One sentence from the bilingual pair (randomly chosen to be either English or Indic).
- `positive`: The aligned translation from the pair.
- `negative`: A randomly sampled sentence from the same language as the anchor, but semantically unrelated.
- **Step 3: Column Renaming & Structuring**
- Original columns like `sentence_en` and `sentence_hi` were renamed to `anchor` and `positive` based on directionality.
- Negative samples were injected from a shuffled pool and assigned to the `negative` column.
- **Step 4: Directionality Randomization**
To avoid bias, each triplet randomly flips between Indic→English and English→Indic.
---
### 📦 Dataset Format
- File type: `.jsonl`
- Each line contains:
```json
{
"anchor": "मैं स्कूल जा रहा हूँ।",
"positive": "I am going to school.",
"negative": "The weather is nice today."
}
```
- Total examples: 60,000
- Languages: Hindi, Bengali, Tamil, Marathi, Gujarati, Punjabi, Kannada, Malayalam, Oriya, Assamese, Telugu, English
---
### 🎯 Intended Use
- Fine-tuning multilingual embedding models (e.g., Gemma, BGE, LaBSE)
- Training contrastive or triplet loss models
- Cross-lingual semantic retrieval
- Evaluation of embedding alignment across languages
---
### 🧪 Supported Tasks
| Task | Description |
|--------------------------|--------------------------------------------------|
| Sentence Embedding | Learn language-agnostic representations |
| Semantic Similarity | Evaluate cosine similarity between anchor/positive |
| Cross-lingual Retrieval | Retrieve aligned sentences across languages |
| Contrastive Learning | Train models to distinguish semantic similarity |
---
### ⚖️ Language Balance
Each language contributes ~5,454 triplets, ensuring balanced representation. Directionality is randomized to prevent source-target bias.
---
### 🔐 License
- License: CC-BY 4.0 (inherits from Samanantar)
- Free for academic, commercial, and open-source use
- Attribution required
---
### 🛠 Preprocessing Tips
- Tokenize using model-specific tokenizer (e.g., GemmaTokenizer)
- Truncate or chunk long sequences to fit model limits
- Optional: Add language tags for anchor/positive/negative for analysis
---
### 📈 Evaluation Metrics
- Cosine similarity
- Mean Reciprocal Rank (MRR)
- nDCG
- Retrieval accuracy
---
### 👤 Maintainer
- **Author**: Parvesh Rawal (XenArcAI)
--- |