URNOVA is a compact Urdu masked-language model encoder built with the ModernBERT architecture.
It was trained from scratch on an Urdu corpus derived from HPLT 3.0 and is designed for:
- Urdu representation learning
- masked-language modelling
- semantic feature extraction
- document and sentence classification
- downstream Urdu NLP fine-tuning
- research involving longer Urdu sequences
Benchmark results
UrBLiMP zero-shot linguistic evaluation
Zero-shot minimal-pair evaluation on UrBLiMP using left-to-right word pseudo-log-likelihood (pll_word_l2r).
| Model | Params | Pairs | Micro Acc. | Macro | Aspect Agreement | Dative Object | Ergativity | Experiencer Subject | Honorific | Noun Phrase | Oblique | Participial Relatives | Subject-Verb Agreement | Word Order |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| URNOVA-95M | 95.2M | 5,698 | 98.17% | 97.96% | 100.00% | 96.23% | 99.10% | 98.77% | 97.39% | 99.50% | 98.22% | 94.68% | 96.66% | 99.01% |
| HPLT-BERT Urdu | 150.2M | 5,698 | 98.48% | 98.14% | 99.13% | 99.14% | 99.28% | 97.04% | 96.41% | 97.75% | 98.54% | 98.01% | 97.62% | 98.51% |
| XLM-R-large | 560.1M | 5,698 | 95.75% | 94.31% | 99.25% | 96.23% | 96.20% | 95.31% | 79.74% | 93.00% | 96.59% | 92.36% | 96.40% | 98.02% |
| Urdu-RoBERTa-small | 126.0M | 5,698 | 94.02% | 90.75% | 99.75% | 88.70% | 97.05% | 95.80% | 70.59% | 92.50% | 96.23% | 89.70% | 89.07% | 88.12% |
| XLM-R-base | 278.3M | 5,698 | 92.70% | 89.32% | 99.50% | 93.84% | 95.36% | 92.84% | 61.44% | 88.50% | 96.52% | 79.40% | 87.79% | 98.02% |
| mmBERT-small | 140.9M | 5,698 | 91.08% | 88.33% | 99.25% | 91.78% | 95.06% | 73.83% | 65.36% | 87.00% | 92.89% | 89.70% | 89.46% | 99.01% |
Quick start
Fill-mask inference
from transformers import pipeline
fill_mask = pipeline(
task="fill-mask",
model="ProximaAI/urnova-95m",
token=True,
)
results = fill_mask(
"پاکستان کی قومی زبان [MASK] ہے۔",
top_k=5,
)
for result in results:
print(
result["token_str"],
result["score"],
)
Load the masked-language model
from transformers import (
AutoModelForMaskedLM,
AutoTokenizer,
)
model_id = "ProximaAI/urnova-95m"
tokenizer = AutoTokenizer.from_pretrained(
model_id,
token=True,
)
model = AutoModelForMaskedLM.from_pretrained(
model_id,
token=True,
)
Extract contextual representations
import torch
from transformers import AutoModel, AutoTokenizer
model_id = "ProximaAI/urnova-95m"
tokenizer = AutoTokenizer.from_pretrained(
model_id,
token=True,
)
encoder = AutoModel.from_pretrained(
model_id,
token=True,
)
text = "اردو ایک خوب صورت اور وسیع زبان ہے۔"
inputs = tokenizer(
text,
return_tensors="pt",
truncation=True,
max_length=8192,
)
with torch.inference_mode():
output = encoder(**inputs)
token_embeddings = output.last_hidden_state
attention_mask = inputs["attention_mask"].unsqueeze(-1)
sentence_embedding = (
token_embeddings * attention_mask
).sum(dim=1) / attention_mask.sum(dim=1)
Training data and provenance
URNOVA was trained from scratch using an Urdu corpus derived from HPLT 3.0.
Recorded corpus details:
- 1,921,027 source documents
- 1,246,973,570 source tokens
- approximately 4 billion total training tokens
- tokenizer vocabulary size of 50,048
- no third-party model initialization
No raw training corpus is distributed in this repository.
See:
Intended use
URNOVA is intended for:
- Urdu NLP research
- masked-language modelling
- encoder-based feature extraction
- document classification
- sentence classification
- semantic similarity research
- downstream supervised fine-tuning
- experimentation with long Urdu documents
Limitations
- The reported benchmark evaluates grammatical minimal-pair preference on UrBLiMP and should not be interpreted as a complete measure of general Urdu language understanding.
- Benchmark performance may not generalize to Roman Urdu, code-switched text, regional dialects, or unrelated domains.
- Minimal-pair pseudo-log-likelihood evaluation measures sentence preference rather than generative ability, reasoning ability, or downstream task performance.
- UrBLiMP results depend on the evaluation and scoring procedure described above and should only be compared with results produced using a compatible methodology.
- Long-context capability does not guarantee reliable understanding over every 8,192-token input.
- Training on web-derived text may reproduce factual errors, stereotypes, offensive content, or representation gaps found in the source material.
Responsible use
Predictions from URNOVA should not be the sole basis for:
- punitive decisions
- political profiling
- content removal
- employment decisions
- legal decisions
- eligibility decisions
- other high-impact automated actions
Applications should include domain-specific validation, human review, and appropriate error analysis.
License and data notice
The original URNOVA model weights, tokenizer artifacts, code, documentation, benchmark reports, and visualizations are made available under the Apache License 2.0, to the extent of the rights held by their respective copyright holders.
The model was trained using an Urdu corpus derived from HPLT 3.0. HPLT licenses its dataset packaging under CC0 but states that it does not own the underlying extracted web text.
Apache-2.0 therefore does not grant rights in underlying training documents or third-party text that may be reproduced by the model.
Third-party benchmark datasets are not covered by URNOVA's Apache-2.0 license. UrBLiMP remains subject to its upstream MIT License and copyright notice.