---
library_name: peft
tags:
- generated_from_trainer
base_model: microsoft/phi-2
model-index:
- name: phi-2-universal-NER
  results: []
datasets:
- Universal-NER/Pile-NER-type
language:
- en
---

# phi-2-universal-NER

This model is a fine-tuned version of [microsoft/phi-2](https://huggingface.co/microsoft/phi-2) on the Universal-NER/Pile-NER-type dataset.

## Model description

This model shows power of small language model. We can finetune phi-2 on google colab free version. It's very simple and easy. I couldn't fine tuned whole model on free colab so used PEFT.

## Intended uses & limitations

This model is fine tuned from Phi-2 and UniversalNER dataset. 

Phi-2 model license changed to MIT but UniversalNER is still under research license so this model can be used for research purpose only.

## Training and evaluation data

I have used just 5 epochs in fine tuning. 

## Training procedure notebook

https://github.com/mit1280/fined-tuning/blob/main/phi_2_fine_tune_using_PEFT%2Binference.ipynb

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0002
- train_batch_size: 2
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- training_steps: 1000

### Inference Code

```python
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
from transformers import StoppingCriteria

config = PeftConfig.from_pretrained("Mit1208/phi-2-universal-NER")
base_model = AutoModelForCausalLM.from_pretrained("microsoft/phi-2",device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(base_model, "Mit1208/phi-2-universal-NER", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("Mit1208/phi-2-universal-NER", trust_remote_code=True)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

conversations = [ { "from": "human", "value": "Text: Mit Patel here from India"}, {"from": "gpt", "value": "I've read this text."}, 
                   {"from":"human", "value":"what is a name of the person in the text?"}]
inference_text = tokenizer.apply_chat_template(conversations, tokenize=False) + '<|im_start|>gpt:\n'
inputs = tokenizer(inference_text, return_tensors="pt", return_attention_mask=False).to(device)

class EosListStoppingCriteria(StoppingCriteria):
    def __init__(self, eos_sequence = tokenizer.encode("<|im_end|>")):
        self.eos_sequence = eos_sequence

    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        last_ids = input_ids[:,-len(self.eos_sequence):].tolist()
        return self.eos_sequence in last_ids

outputs = model.generate(**inputs, max_length=512, pad_token_id= tokenizer.eos_token_id,
            stopping_criteria = [EosListStoppingCriteria()])

text = tokenizer.batch_decode(outputs)[0]

print(text)

# Output
'''
<|im_start|>human
Text: Mit Patel here from India<|im_end|>
<|im_start|>gpt
I've read this text.<|im_end|>
<|im_start|>human
what is a name of the person in the text?<|im_end|>
<|im_start|>gpt:
["Mit Patel"]<|im_end|>
'''
```


### Framework versions

- PEFT 0.7.1
- Transformers 4.36.2
- Pytorch 2.1.0+cu121
- Datasets 2.15.0
- Tokenizers 0.15.0