---
language: tr
library_name: peft
pipeline_tag: token-classification
base_model: dbmdz/bert-base-turkish-cased
---
## Training procedure

This is a fine-tuned model of base model "dbmdz/bert-base-turkish-cased" using the Parameter Efficient Fine Tuning (PEFT) with Low-Rank Adaptation (LoRA) technique 
using a reviewed version of well known Turkish NER dataset 
(https://github.com/stefan-it/turkish-bert/files/4558187/nerdata.txt).

trainable params: 702,734 || all params: 110,627,342 || trainable%: 0.6352263258752072

# Fine-tuning parameters:
```
task = "ner"
model_checkpoint = "dbmdz/bert-base-turkish-cased"
batch_size = 16 
label_list = ['O', 'B-PER', 'I-PER', 'B-ORG', 'I-ORG', 'B-LOC', 'I-LOC']
max_length = 512 
learning_rate = 1e-3 
num_train_epochs = 7 
weight_decay = 0.01
```
# PEFT Parameters
```
inference_mode=False
r=16
lora_alpha=16
lora_dropout=0.1
bias="all"
```
# How to use: 
```
peft_model_id = "akdeniz27/bert-base-turkish-cased-ner-lora"
config = PeftConfig.from_pretrained(peft_model_id)
inference_model = AutoModelForTokenClassification.from_pretrained(
    config.base_model_name_or_path, num_labels=7, id2label=id2label, label2id=label2id
)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
model = PeftModel.from_pretrained(inference_model, peft_model_id)
text = "Mustafa Kemal Atatürk 1919 yılında Samsun'a çıktı."
inputs = tokenizer(text, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
tokens = inputs.tokens()
predictions = torch.argmax(logits, dim=2)
for token, prediction in zip(tokens, predictions[0].numpy()):
    print((token, model.config.id2label[prediction]))
```

# Reference test results:
* accuracy: 0.993297
* f1: 0.949696
* precision: 0.942554
* recall: 0.956946