---
license: apache-2.0
pipeline_tag: sentence-similarity
library_name: transformers
language:
- en
- ar
- ur
tags:
- embeddings
- sentence-similarity
- retrieval
- contrastive-learning
- multilingual
- from-scratch
datasets:
- sentence-transformers/all-NLI
- facebook/xnli
- Helsinki-NLP/opus-100
- sentence-transformers/msmarco-bm25
- nlpai-lab/miracl-multilingual-triplets
metrics:
- mrr@10
- ndcg@10
- spearman_correlation
- recall@10
---

# mentee-embed-v4

**A 41M-parameter trilingual text embedding model trained entirely from scratch — no pretrained backbone, no BERT, no RoBERTa. Random initialization only.**

v4 adds mMARCO Arabic retrieval data (500K triplets) and a third distillation round with hard negatives, yielding significant improvements over v3 across all languages and benchmarks.

Developed by **Syed Syab Ahmad Shah** and **Team MenteE AI** (menteeai.org) as part of an ongoing research effort to build competitive multilingual embeddings from the ground up for Arabic, English, and Urdu.

---

## Key Facts

| Property | Value |
|----------|-------|
| Parameters | **41M** |
| Embedding dimension | 384 |
| Max sequence length | 128 tokens |
| Languages | Arabic · English · Urdu |
| Architecture | 12-layer Transformer, custom BPE tokenizer (50K vocab) |
| Initialization | **Random** (trained from scratch) |
| Training data | ~2.6M triplets (NLI + MS-MARCO + OPUS + MIRACL + mMARCO Arabic) |
| Pooling | Mean pooling |
| Training objective | Relational distillation + InfoNCE contrastive (teacher: `multilingual-e5-base`) |

---

## What "From Scratch" Means

Most embedding models fine-tune an existing pretrained encoder (BERT, RoBERTa, MPNet). **mentee-embed-v4 does not.** We:

1. Trained a **custom BPE tokenizer** on Arabic, English, and Urdu text
2. Initialized a **12-layer Transformer with random weights**
3. Ran **masked language modeling** pretraining (50K steps on trilingual corpus)
4. Applied **three-round contrastive distillation** with hard negative mining

No pretrained checkpoint was used at any stage.

---

## v4 vs v3 Improvements

| Benchmark | v3 | v4 | Change |
|-----------|----|----|--------|
| Bench MRR@10 (all) | 0.103 | **0.252** | +146% |
| MIRACL EN MRR@10 | 0.768 | **0.916** | +19% |
| MIRACL AR MRR@10 | 0.447 | **0.874** | +96% |
| MIRACL UR MRR@10 | 0.354 | **0.572** | +62% |
| MS-MARCO MRR@10 | 0.592 | **0.706** | +19% |
| STS-B Spearman | 0.683 | **0.725** | +6% |

Key improvements come from:
- **mMARCO Arabic** — 500K Arabic retrieval triplets added to training
- **50K MLM steps** — 6x more pretraining than v3 (8K steps)
- **3 distillation rounds** — extra hard-negative round for better discrimination

---

## Full Benchmark Results

All baselines evaluated under identical conditions on the same hardware (RTX 5090).

### MenteE Embed Bench — Custom EN/AR/UR Pilot

*121 queries · 18 domains · dialect Arabic · Roman Urdu · hard negatives*

| Model | Params | EN MRR@10 | AR MRR@10 | UR MRR@10 | Dialect AR | Roman UR | All MRR@10 |
|---|---|---|---|---|---|---|---|
| **mentee-embed-v4** | **41M** | 0.369 | 0.120 | 0.261 | 0.095 | 0.399 | **0.252** |
| paraphrase-mpnet-base-v2 | 278M | 0.764 | 0.602 | 0.611 | 0.445 | 0.661 | 0.661 |
| paraphrase-MiniLM-L12-v2 | 118M | 0.682 | 0.568 | 0.519 | 0.385 | 0.480 | 0.592 |
| multilingual-e5-base | 278M | 0.667 | 0.352 | 0.413 | 0.310 | 0.430 | 0.481 |
| all-MiniLM-L6-v2 | 23M | 0.873 | 0.087 | 0.289 | 0.084 | 0.486 | 0.425 |

![Custom Bench MRR@10](benchmarks/v4_full/charts/bench_mrr_all.png)

#### Per-Language Breakdown

![Per-Language MRR@10](benchmarks/v4_full/charts/bench_per_lang.png)

| Metric | v4 | mpnet | e5-base | MiniLM-L6 |
|--------|-----|-------|---------|-----------|
| English MRR@10 | 0.369 | 0.764 | 0.667 | 0.873 |
| Arabic MRR@10 | 0.120 | 0.602 | 0.352 | 0.087 |
| Urdu MRR@10 | 0.261 | 0.611 | 0.413 | 0.289 |
| Dialect AR MRR@10 | 0.095 | 0.445 | 0.310 | 0.084 |
| Roman UR MRR@10 | 0.399 | 0.661 | 0.430 | 0.486 |

#### Detailed Metrics (NDCG@10, Recall@1, Recall@5, Recall@10)

| Model | All NDCG | All R@1 | All R@5 | All R@10 | EN NDCG | AR NDCG | UR NDCG |
|---|---|---|---|---|---|---|---|
| **mentee-embed-v4** | 0.238 | 0.083 | 0.203 | 0.273 | 0.331 | 0.166 | 0.212 |
| mpnet-base | 0.650 | 0.438 | 0.620 | 0.736 | 0.757 | 0.589 | 0.595 |
| MiniLM-L12 | 0.580 | 0.355 | 0.562 | 0.661 | 0.674 | 0.556 | 0.504 |
| e5-base | 0.468 | 0.248 | 0.446 | 0.549 | 0.653 | 0.341 | 0.398 |
| MiniLM-L6 | 0.412 | 0.231 | 0.376 | 0.483 | 0.868 | 0.104 | 0.266 |

#### Per-Domain MRR@10 (v4)

![v4 Per-Domain MRR@10](benchmarks/v4_full/charts/v4_per_domain.png)

| Domain | MRR@10 | Domain | MRR@10 |
|---|---|---|---|
| Travel | 0.643 | Fashion | 0.200 |
| Hotels | 0.378 | History | 0.195 |
| Restaurants | 0.357 | Sports | 0.179 |
| Technology | 0.333 | Health | 0.164 |
| Finance | 0.286 | Entertainment | 0.143 |
| Automotive | 0.222 | Social Media | 0.029 |

---

### Protocol A — In-batch Retrieval (MIRACL)

*acc@1 / MRR@10*

| Model | MIRACL-EN | MIRACL-AR | MIRACL-UR |
|-------|-----------|-----------|-----------|
| **mentee-embed-v4** | 0.870 / **0.916** | 0.825 / **0.874** | 0.475 / **0.572** |
| paraphrase-mpnet-base-v2 | 0.965 / 0.982 | 0.860 / 0.898 | 0.755 / 0.824 |
| paraphrase-MiniLM-L12-v2 | 0.965 / 0.979 | 0.815 / 0.868 | 0.720 / 0.785 |
| multilingual-e5-base | 0.980 / 0.990 | 0.940 / 0.958 | 0.960 / 0.970 |
| all-MiniLM-L6-v2 | 0.980 / 0.990 | 0.005 / 0.027 | 0.015 / 0.031 |

![MIRACL Protocol A](benchmarks/v4_full/charts/miracl_protocol_a.png)

> v4 achieves strong MIRACL scores despite being 6.8x smaller than mpnet-base. MIRACL AR (0.874) is within 3% of mpnet-base (0.898).

---

### Protocol C — MS-MARCO Corpus Retrieval (10K passages)

| Model | Params | MRR@10 | R@5 | R@100 |
|-------|--------|--------|-----|-------|
| all-MiniLM-L6-v2 | 23M | 0.877 | 0.992 | 0.992 |
| multilingual-e5-base | 278M | 0.856 | 0.992 | 0.996 |
| paraphrase-mpnet-base-v2 | 278M | 0.803 | 0.964 | 0.988 |
| **mentee-embed-v4** | **41M** | **0.706** | **0.913** | **0.980** |

![MS-MARCO Protocol C](benchmarks/v4_full/charts/msmarco_protocol_c.png)

---

### STS-B (Semantic Textual Similarity)

| Model | Spearman | Pearson | MAE | RMSE |
|-------|----------|---------|-----|------|
| paraphrase-mpnet-base-v2 | 0.868 | 0.870 | — | — |
| bge-small-en-v1.5 | 0.859 | — | — | — |
| paraphrase-MiniLM-L12-v2 | 0.844 | 0.849 | — | — |
| multilingual-e5-base | 0.842 | 0.847 | — | — |
| **mentee-embed-v4** | **0.725** | **0.730** | **2.059** | **2.458** |

![STS-B Spearman](benchmarks/v4_full/charts/stsb_spearman.png)

---

## Speed & Efficiency

Measured on RTX 5090.

| Model | Params | Sents/sec (bs=128) | Latency@1 (ms) | VRAM (MB) | Cost/1B sents |
|---|---|---|---|---|---|
| **mentee-embed-v4** | **41M** | **18,115** | **5.15** | **175** | **$7.15** |
| all-MiniLM-L6-v2 | 23M | 12,445 | 4.71 | 103 | $10.40 |
| multilingual-e5-small | 118M | 9,749 | 7.88 | 482 | $13.28 |
| paraphrase-MiniLM-L12-v2 | 118M | 9,456 | 8.17 | 573 | $13.69 |
| multilingual-e5-base | 278M | 5,379 | 8.07 | 1126 | $24.07 |
| paraphrase-mpnet-base-v2 | 278M | 5,158 | 8.04 | 1126 | $25.10 |

![Speed Comparison](benchmarks/v4_full/charts/speed_comparison.png)

> v4 is the **fastest model tested** — 3.5x faster than mpnet-base, 45% faster than MiniLM-L12.

### Batch Size Scaling

| Batch Size | Sents/sec | Latency (ms) | Peak VRAM (MB) |
|---|---|---|---|
| 1 | 145 | 6.90 | 174 |
| 8 | 902 | 8.87 | 177 |
| 16 | 1,707 | 9.37 | 181 |
| 32 | 3,424 | 9.35 | 188 |
| 64 | 6,470 | 9.89 | 202 |
| 128 | 12,418 | 10.31 | 230 |
| 256 | 20,546 | 12.46 | 287 |
| 512 | 22,139 | 23.13 | 400 |

### Precision Formats

| Format | Sents/sec (bs=128) | STS-B Spearman | Delta vs FP32 | VRAM (MB) |
|---|---|---|---|---|
| FP32 | 12,500 | 0.7249 | — | 289 |
| FP16 | 3,632 | 0.7249 | 0.0 | 149 |
| BF16 | 2,837 | 0.7248 | -0.0001 | 149 |

### Efficiency Metrics (Per Parameter)

| Model | Bench MRR/1M params | Sents/sec per 1M params | Cost per 1B sents |
|---|---|---|---|
| **mentee-embed-v4** | **0.525** | **441.82** | **$7.15** |
| all-MiniLM-L6-v2 | 0.344 | 541.09 | $10.40 |
| paraphrase-MiniLM-L12-v2 | 0.452 | 80.34 | $13.69 |
| multilingual-e5-base | 0.166 | 19.35 | $24.07 |
| paraphrase-mpnet-base-v2 | 0.224 | 18.55 | $25.10 |

![Efficiency](benchmarks/v4_full/charts/efficiency.png)
![Cost Comparison](benchmarks/v4_full/charts/cost_comparison.png)

---

## Radar — v4 vs mpnet-base

Normalized comparison across all evaluation dimensions:

![Radar — v4 vs mpnet](benchmarks/v4_full/charts/v4_radar.png)

| Dimension | v4 (41M) | mpnet (278M) | v4 as % of mpnet |
|---|---|---|---|
| English MRR@10 | 0.369 | 0.764 | 48% |
| Arabic MRR@10 | 0.120 | 0.602 | 20% |
| Urdu MRR@10 | 0.261 | 0.611 | 43% |
| MIRACL EN | 0.916 | 0.982 | 93% |
| MIRACL AR | 0.874 | 0.898 | 97% |
| MIRACL UR | 0.572 | 0.824 | 69% |
| MS-MARCO | 0.706 | 0.803 | 88% |
| STS-B Spearman | 0.725 | 0.868 | 84% |
| Throughput | 18,115 | 5,158 | 351% |
| Cost efficiency | $7.15 | $25.10 | 351% |

---

## Honest Assessment

### Where v4 Wins
- **Speed** — fastest embedding model tested (18,115 sents/sec)
- **Cost** — cheapest per billion sentences ($7.15)
- **MIRACL AR** — 0.874 vs mpnet's 0.898 (within 3%) despite 6.8x fewer params
- **MS-MARCO** — 0.706, competitive with models 3-7x larger
- **Efficiency** — best MRR per parameter of all models tested

### Where v4 Loses
- **Custom bench** — 0.252 vs mpnet's 0.661 (2.4x gap)
- **English** — 0.369 vs MiniLM-L6's 0.873
- **STS-B** — 0.725 vs mpnet's 0.868 (16% gap)
- **Arabic dialect** — 0.095 (struggles with informal Arabic)
- **Urdu** — 0.261 vs mpnet's 0.611

### Verdict

v4 is a **strong proof-of-concept** — it proves you can train a competitive embedding model from scratch. The MIRACL AR score (0.874) is within 3% of mpnet-base (0.898) despite being 6.8x smaller.

For production use, v4 works well for:
- Arabic retrieval (MIRACL-scale tasks)
- High-throughput, cost-sensitive applications
- Resource-constrained environments (175 MB VRAM)

v4 is not yet suitable for:
- Domain-specific English retrieval
- Dialectal Arabic / informal Urdu
- Tasks requiring STS-level semantic similarity

---

## Training Data

| Source | Language | Triplets | Type |
|--------|----------|----------|------|
| all-NLI (sentence-transformers) | EN | 558K | NLI triplets |
| XNLI | AR | 128K | NLI triplets |
| XNLI | UR | 125K | NLI triplets |
| OPUS-100 EN-UR | EN/UR | 300K | Parallel translation |
| OPUS-100 AR-EN | AR/EN | 300K | Parallel translation |
| MS-MARCO BM25 triplets | EN | 500K | Passage retrieval |
| MS-MARCO hard negatives | EN | 200K | Hard retrieval |
| mMARCO Arabic | AR | 500K | Passage retrieval |
| MIRACL | EN/AR/UR | ~9K | Wikipedia retrieval |
| **Total** | | **~2.6M** | |

---

## Training Pipeline

```
Stage 1 — MLM Pretraining (50K steps)
  Random init → masked language modeling on trilingual corpus
  50,000 steps · batch=64 · vocab=50K BPE

Stage 2 — Teacher Encoding
  Teacher: intfloat/multilingual-e5-base (768-dim)
  Encode all 2.6M triplets to soft labels

Stage 3 — Distillation Round 1 (no hard negatives)
  InfoNCE contrastive + relational distillation
  4,000 steps · batch=512 · temp=0.05

Stage 4 — Hard Negative Mining Round 1
  GPU-accelerated top-5 mining across full 2.6M corpus

Stage 5 — Distillation Round 2 (with hard negatives)
  Same objective + mined hard negatives per anchor
  10,000 steps · batch=512 · temp=0.05

Stage 6 — Hard Negative Mining Round 2 (on improved model)
  Re-mine with the better round-2 model

Stage 7 — Distillation Round 3 (final polish)
  Train on round-2 mined negatives
  15,000 cumulative steps · batch=512 · temp=0.05
```

---

## Usage

```python
# pip install torch transformers tokenizers huggingface_hub
from transformers import AutoModel, AutoTokenizer

tok   = AutoTokenizer.from_pretrained("MenteEAI/mentee-embed-v4", trust_remote_code=True)
model = AutoModel.from_pretrained("MenteEAI/mentee-embed-v4",   trust_remote_code=True)

sentences = [
    "Hello, how are you?",
    "مرحبا، كيف حالك؟",
    "ہیلو، آپ کیسے ہیں؟"
]

embeddings = model.encode(sentences, tokenizer=tok)
print(embeddings.shape)  # torch.Size([3, 384])
```

> `trust_remote_code=True` is required — standard for custom-architecture models on HuggingFace. The code runs entirely on your machine.

### Similarity search
```python
from transformers import AutoModel, AutoTokenizer

tok   = AutoTokenizer.from_pretrained("MenteEAI/mentee-embed-v4", trust_remote_code=True)
model = AutoModel.from_pretrained("MenteEAI/mentee-embed-v4",   trust_remote_code=True)

query    = model.encode(["What is machine learning?"], tokenizer=tok)
passages = model.encode([
    "Machine learning is a subset of artificial intelligence.",
    "The weather today is sunny.",
    "تعلم الآلة هو فرع من فروع الذكاء الاصطناعي.",
], tokenizer=tok)

scores = query @ passages.T
print(scores)  # tensor([[0.81, 0.60, 0.79]])
```

---

## Limitations

- **Custom benchmark gap** — v4 scores 0.252 vs 0.661 for mpnet-base on the custom bench; the gap reflects the difference between 2.6M training pairs and web-scale data
- **Arabic and Urdu** lag behind English on the custom bench, though MIRACL scores are strong (AR 0.874, UR 0.572)
- Vocabulary limited to 50K tokens trained on ~2.6M sentences — rare scripts and dialects may tokenize poorly
- Not evaluated on MTEB full suite yet

---

## Citation

```bibtex
@misc{mentee-embed-v4-2026,
  title   = {How Far Can Multilingual Text Embeddings Be Trained From Scratch?
             A Compute-Efficient Study of Arabic, English, and Urdu},
  author  = {Shah, Syed Syab Ahmad and Team MenteE AI},
  year    = {2026},
  url     = {https://huggingface.co/MenteEAI/mentee-embed-v4},
  note    = {MenteE AI. Apache-2.0 License}
}
```

---

## About MenteE AI

Built by **Syed Syab Ahmad Shah** and Team MenteE AI.
[menteeai.org](https://menteeai.org) · syab@menteeai.org