kallamni-embed-v1 — Emirati Spoken Arabic Embedding Model

Author: @yasserrmd
Version: v1
License: Apache 2.0

🎯 Motivation

kallamni-embed-v1 was built to address a gap in Arabic NLP — the absence of a high-fidelity model for spoken Emirati Arabic.
While most Arabic embeddings (AraBERT, CAMeLBERT, MARBERT) focus on MSA or pan-Arab dialects, they fail to capture UAE’s informal patterns such as:

Lexical variants: وايد, مب, سير, ويّاكم
Code-switching: “bro yalla lets go al mall”
Arabizi + emojis: “ana mb 3arf 😅 sho y9eer!”

This model learns these naturally occurring forms using curated Emirati-style Q&A and conversation datasets.

SentenceTransformer based on BAAI/bge-m3

This is a sentence-transformers model finetuned from BAAI/bge-m3. It maps sentences & paragraphs to a 1024-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

Model Type: Sentence Transformer
Base model: BAAI/bge-m3
Maximum Sequence Length: 8192 tokens
Output Dimensionality: 1024 dimensions
Similarity Function: Cosine Similarity

Model Sources

Documentation: Sentence Transformers Documentation
Repository: Sentence Transformers on GitHub
Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 8192, 'do_lower_case': False}) with Transformer model: XLMRobertaModel 
  (1): Pooling({'word_embedding_dimension': 1024, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  (2): Normalize()
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("sentence_transformers_model_id")
# Run inference
sentences = [
    'كيف كانت التجربة في المطعم اليديد؟',
    'المطعم كان ممتاز، الأكل لذيذ والخدمة سريعة.',
    'كنت وايد سعيد، السوالف ما خلصت بيننا.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1024]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

Training Details

Training Dataset

Unnamed Dataset

Size: 50,000 training samples
Columns: sentence_0 and sentence_1
Approximate statistics based on the first 1000 samples:
sentence_0 sentence_1
type string string
details
min: 7 tokens
mean: 13.47 tokens
max: 24 tokens

min: 8 tokens
mean: 18.85 tokens
max: 36 tokens

	sentence_0	sentence_1
type	string	string
details	min: 7 tokens mean: 13.47 tokens max: 24 tokens	min: 8 tokens mean: 18.85 tokens max: 36 tokens

Samples:

sentence_0	sentence_1
`قد استخدمت تطبيق تتبع السعرات الحرارية؟`	`إيه، يساعدني في مراقبة أكلي ونسبة البروتين.`
`شو كانت أول تجربة لك في التدريب العملي؟`	`كانت مميزة، استفدت وتعلمت أشياء ما تدرسها الكتب.`
`إذا حد قال 'على عينه حار'، شو يقصد؟`	`يعني هذا شخص صريح وما يجامل، يقول اللي في قلبه.`

Loss: MultipleNegativesRankingLoss with these parameters:

{
    "scale": 20.0,
    "similarity_fct": "cos_sim"
}

Training Hyperparameters

Non-Default Hyperparameters

per_device_train_batch_size: 24
per_device_eval_batch_size: 24
fp16: True
multi_dataset_batch_sampler: round_robin

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: no
prediction_loss_only: True
per_device_train_batch_size: 24
per_device_eval_batch_size: 24
per_gpu_train_batch_size: None
per_gpu_eval_batch_size: None
gradient_accumulation_steps: 1
eval_accumulation_steps: None
torch_empty_cache_steps: None
learning_rate: 5e-05
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1
num_train_epochs: 3
max_steps: -1
lr_scheduler_type: linear
lr_scheduler_kwargs: {}
warmup_ratio: 0.0
warmup_steps: 0
log_level: passive
log_level_replica: warning
log_on_each_node: True
logging_nan_inf_filter: True
save_safetensors: True
save_on_each_node: False
save_only_model: False
restore_callback_states_from_checkpoint: False
no_cuda: False
use_cpu: False
use_mps_device: False
seed: 42
data_seed: None
jit_mode_eval: False
use_ipex: False
bf16: False
fp16: True
fp16_opt_level: O1
half_precision_backend: auto
bf16_full_eval: False
fp16_full_eval: False
tf32: None
local_rank: 1
ddp_backend: None
tpu_num_cores: None
tpu_metrics_debug: False
debug: []
dataloader_drop_last: True
dataloader_num_workers: 0
dataloader_prefetch_factor: None
past_index: -1
disable_tqdm: False
remove_unused_columns: True
label_names: None
load_best_model_at_end: False
ignore_data_skip: False
fsdp: []
fsdp_min_num_params: 0
fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
fsdp_transformer_layer_cls_to_wrap: None
accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
deepspeed: None
label_smoothing_factor: 0.0
optim: adamw_torch
optim_args: None
adafactor: False
group_by_length: False
length_column_name: length
ddp_find_unused_parameters: None
ddp_bucket_cap_mb: None
ddp_broadcast_buffers: False
dataloader_pin_memory: True
dataloader_persistent_workers: False
skip_memory_metrics: True
use_legacy_prediction_loop: False
push_to_hub: False
resume_from_checkpoint: None
hub_model_id: None
hub_strategy: every_save
hub_private_repo: None
hub_always_push: False
gradient_checkpointing: False
gradient_checkpointing_kwargs: None
include_inputs_for_metrics: False
include_for_metrics: []
eval_do_concat_batches: True
fp16_backend: auto
push_to_hub_model_id: None
push_to_hub_organization: None
mp_parameters:
auto_find_batch_size: False
full_determinism: False
torchdynamo: None
ray_scope: last
ddp_timeout: 1800
torch_compile: False
torch_compile_backend: None
torch_compile_mode: None
include_tokens_per_second: False
include_num_input_tokens_seen: False
neftune_noise_alpha: None
optim_target_modules: None
batch_eval_metrics: False
eval_on_start: False
use_liger_kernel: False
eval_use_gather_object: False
average_tokens_across_devices: False
prompts: None
batch_sampler: batch_sampler
multi_dataset_batch_sampler: round_robin

Training Logs

Epoch	Step	Training Loss
0.4803	500	0.3377
0.9606	1000	0.1394
1.4409	1500	0.0828
1.9212	2000	0.0465
2.4015	2500	0.0317
2.8818	3000	0.0211

Evaluation Overview

V4 — Hyper-Authentic Emirati Benchmark

Metric	multilingual-e5-large	kallamni-embed-v1
nDCG@10	0.0268	0.0421
MRR	0.0322	0.0437
Precision@1	0.0133	0.0267
Pearson Corr	−0.2718	−0.0963
F1	1.000	1.000

→ +57 % gain in retrieval relevance over the multilingual baseline.

V5 — Dialect Robustness Benchmark

Subset	multilingual-e5-large	kallamni-embed-v1
PURE EMI	0.0359	0.0582
ARABIZI + EMOJI	0.0012	0.0167
CODE-SWITCH	0.0010	0.0219
GULF OTHER	0.0543	0.0469
SOCIAL NOISE	0.0127	0.0334
CONTROL MIX	0.0157	0.0386

Statistical significance: Δ nDCG@10 = +0.0218 (95 % CI [0.0008 – 0.0439], p = 0.04)

📈 Visual Summary

The Emirati-tuned model maintains high stability across dialectal noise — especially Arabizi, Code-Switch, and Social Noise subsets — where multilingual models collapse.

🧠 Robustness & Use Cases

Handles informal input: Arabizi, emojis, typos, and Gulf-accented syntax.
Optimized for retrieval & RAG: Works well in vector databases for Emirati chatbots, citizen-service platforms, and multilingual UAE apps.
Fast inference: ~15 % faster than multilingual-e5-large on average batch size 32.
Cross-dialect adaptability: Maintains coherence on Gulf-neighbor variations (Kuwaiti, Omani).

🧩 Why Other Models Were Excluded

Model	nDCG@10 (pilot)	Pearson	Comment
CAMeLBERT-DA	0.018	−0.42	Trained on MSA + Levantine Twitter, weak Emirati signal
AraBERT v2	0.023	−0.38	Diacritic bias, poor slang handling
MARBERT	0.031	−0.29	Broad Gulf coverage, low UAE lexical overlap
mE5-base	0.025	−0.31	Generic multilingual, not dialect-aware

These models were retained for reference but excluded from the final leaderboard because they lack UAE-specific conversational grounding.

🔬 Benchmark Protocol

All datasets were auto-synthesized inside the evaluation script to ensure control and reproducibility.

Retrieval pairs: 500 queries × 500 docs (3 hard negatives per gold)
Similarity pairs: 2 000 sentence pairs
Classification: 3 600 texts across 3 classes (Complaint / Humor / Question)
5-fold cross-validation + paired bootstrap CIs

Intended Use

Task	Description	Example
Semantic Search	Embed Emirati chat data for retrieval	“وين المكان اللي في الصورة؟” → relevant caption
Conversational RAG	Retrieve contextually similar utterances	“شو معنى كلمة مب؟”
Intent Classification	Complaint vs Informal chat vs Inquiry	“السيارة ما تشتغل من أمس 😡”

Framework Versions

Python: 3.11.13
Sentence Transformers: 4.1.0
Transformers: 4.52.4
PyTorch: 2.6.0+cu124
Accelerate: 1.8.1
Datasets: 3.6.0
Tokenizers: 0.21.2

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

MultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

Downloads last month: 17

Safetensors

Model size

0.6B params

Tensor type

F32

Model tree for yasserrmd/kallamni-embed-v1

Base model

BAAI/bge-m3

Finetuned

(348)

this model

Quantizations

1 model