library_name: transformers
language:
- uz
license: mit
base_model: FacebookAI/xlm-roberta-large
tags:
- generated_from_trainer
datasets:
- risqaliyevds/uzbek_ner
metrics:
- precision
- recall
- f1
- accuracy
model-index:
- name: Uzbek NER model
results: []
Uzbek NER model
This model is a fine-tuned version of FacebookAI/xlm-roberta-large on the Uzbek Ner dataset. It achieves the following results on the evaluation set:
- Loss: 0.1754
- Precision: 0.5848
- Recall: 0.6313
- F1: 0.6071
- Accuracy: 0.9386
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 8
- total_train_batch_size: 64
- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine_with_restarts
- lr_scheduler_warmup_ratio: 0.08
- num_epochs: 3
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Precision | Recall | F1 | Accuracy |
|---|---|---|---|---|---|---|---|
| 0.2474 | 0.4662 | 100 | 0.2283 | 0.4911 | 0.5164 | 0.5035 | 0.9284 |
| 0.2039 | 0.9324 | 200 | 0.1942 | 0.5495 | 0.5836 | 0.5661 | 0.9345 |
| 0.1949 | 1.3963 | 300 | 0.1855 | 0.5591 | 0.6348 | 0.5945 | 0.9359 |
| 0.19 | 1.8625 | 400 | 0.1800 | 0.5604 | 0.6279 | 0.5922 | 0.9361 |
| 0.1769 | 2.3263 | 500 | 0.1761 | 0.5806 | 0.6262 | 0.6025 | 0.9381 |
| 0.1765 | 2.7925 | 600 | 0.1754 | 0.5849 | 0.6311 | 0.6071 | 0.9386 |
Framework versions
- Transformers 4.49.0
- Pytorch 2.5.1+cu124
- Datasets 3.3.2
- Tokenizers 0.21.0
Usage
from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch
custom_id2label = { 0: "O", 1: "B-CARDINAL", 2: "I-CARDINAL", 3: "B-DATE", 4: "I-DATE", 5: "B-EVENT", 6: "I-EVENT", 7: "B-GPE", 8: "I-GPE", 9: "B-LOC", 10: "I-LOC", 11: "B-MONEY", 12: "I-MONEY", 13: "B-ORDINAL", 14: "B-ORG", 15: "I-ORG", 16: "B-PERCENT", 17: "I-PERCENT", 18: "B-PERSON", 19: "I-PERSON", 20: "B-TIME", 21: "I-TIME" }
custom_label2id = {v: k for k, v in custom_id2label.items()}
model_name = "mustafoyev202/roberta-uz"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name, num_labels=23)
model.config.id2label = custom_id2label
model.config.label2id = custom_label2id
text = "Tesla kompaniyasi AQSHda joylashgan."
tokens = tokenizer(text.split(), return_tensors="pt", is_split_into_words=True)
with torch.no_grad(): logits = model(**tokens).logits
predicted_token_class_ids = logits.argmax(-1).squeeze().tolist()
word_ids = tokens.word_ids()
previous_word_id = None
word_predictions = {}
for i, word_id in enumerate(word_ids): if word_id is not None: label = custom_id2label[predicted_token_class_ids[i]] if word_id != previous_word_id: # New word word_predictions[word_id] = label previous_word_id = word_id
words = text.split() # Splitting for simplicity
final_predictions = [(word, word_predictions.get(i, "O")) for i, word in enumerate(words)]
print("Predictions:")
for word, label in final_predictions: print(f"{word}: {label}")
labels = torch.tensor([predicted_token_class_ids]).unsqueeze(0) # Adjust dimensions
loss = model(**tokens, labels=labels).loss
print("\nLoss:", round(loss.item(), 2))