---
language:
- ur
library_name: transformers
pipeline_tag: feature-extraction
tags:
- BERT
- encoder
- embeddings
- TiME
- ur
- size:m
license: apache-2.0
teacher_model: FacebookAI/xlm-roberta-large
datasets:
- uonlp/CulturaX
---

# TiME Urdu (ur, m)

Monolingual BERT-style encoder that outputs embeddings for Urdu.
Distilled from FacebookAI/xlm-roberta-large.

## Specs
- language: Urdu (ur)
- size: m
- architecture: BERT encoder
- layers: 6
- hidden size: 768
- intermediate size: 3072

## Usage (mean pooled embeddings)

```python
from transformers import AutoTokenizer, AutoModel
import torch

repo = "dschulmeist/TiME-ur-m"
tok = AutoTokenizer.from_pretrained(repo)
mdl = AutoModel.from_pretrained(repo)

def mean_pool(last_hidden_state, attention_mask):
    mask = attention_mask.unsqueeze(-1).type_as(last_hidden_state)
    return (last_hidden_state * mask).sum(1) / mask.sum(1).clamp(min=1e-9)

inputs = tok(["example sentence"], padding=True, truncation=True, return_tensors="pt")
outputs = mdl(**inputs)
emb = mean_pool(outputs.last_hidden_state, inputs['attention_mask'])
print(emb.shape)
```