|
--- |
|
license: apache-2.0 |
|
tags: |
|
- generated_from_trainer |
|
- t5 |
|
- flan |
|
- small |
|
- peft |
|
- QLoRA |
|
- cnn_dailymail |
|
datasets: |
|
- cnn_dailymail |
|
model-index: |
|
- name: QLoRA-Flan-T5-Small |
|
results: [] |
|
metrics: |
|
- rouge |
|
--- |
|
|
|
<!-- This model card has been generated automatically according to the information the Trainer had access to. You |
|
should probably proofread and complete it, then remove this comment. --> |
|
|
|
# QLoRA-Flan-T5-Small |
|
|
|
This model is a fine-tuned version of [google/flan-t5-small](https://huggingface.co/google/flan-t5-small) on the cnn_dailymail dataset. It achieves the following on the test set: |
|
|
|
- ROUGE-1: 0.3484265780526604 |
|
- ROUGE-2: 0.14343059577230782 |
|
- ROUGE-l: 0.32809541498574013 |
|
|
|
## Model description |
|
|
|
This model was fine-tuned with the purpose of performing the task of abstractive summarization. |
|
|
|
|
|
## Training and evaluation data |
|
|
|
Fine-tuned on cnn_dailymail training set |
|
Evaluated on cnn_dailymail test set |
|
|
|
## How to use model |
|
|
|
1. Loading the model |
|
|
|
```python |
|
import torch |
|
from peft import PeftModel, PeftConfig |
|
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer |
|
|
|
# Load peft config for pre-trained checkpoint etc. |
|
peft_model_id = "emonty777/QLoRA-Flan-T5-Small" |
|
|
|
config = PeftConfig.from_pretrained(peft_model_id) |
|
|
|
# load base LLM model and tokenizer / runs on CPU |
|
model = AutoModelForSeq2SeqLM.from_pretrained(config.base_model_name_or_path) |
|
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path) |
|
|
|
# load base LLM model and tokenizer for GPU |
|
model = AutoModelForSeq2SeqLM.from_pretrained(config.base_model_name_or_path, load_in_8bit=True, device_map={"":0}) |
|
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path) |
|
|
|
# Load the Lora model |
|
model = PeftModel.from_pretrained(model, peft_model_id, device_map={"":0}) |
|
model.eval() |
|
``` |
|
2. Generating summaries |
|
|
|
```python |
|
text = "Your text goes here..." |
|
|
|
# If you want to use CPU |
|
input_ids = tokenizer(text, return_tensors="pt", truncation=True).input_ids |
|
# If you want to use GPU |
|
input_ids = tokenizer(text, return_tensors="pt", truncation=True).input_ids.cuda() |
|
# Adjust max_new_tokens based on size. This is set up for articles of text |
|
outputs = model.generate(input_ids=input_ids, max_new_tokens=120, do_sample=False) |
|
|
|
print(f"input sentence: {sample['article']}\n{'---'* 20}") |
|
print(f"summary:\n{tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True)[0]}") |
|
|
|
``` |
|
|
|
|
|
## Training procedure |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 3e-05 |
|
- train_batch_size: 8 |
|
- eval_batch_size: 8 |
|
- seed: 42 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: linear |
|
- num_epochs: 4 |
|
|
|
### Training results |
|
|
|
Evaluated on full CNN Dailymail test set |
|
|
|
- ROUGE-1: 0.3484265780526604 |
|
- ROUGE-2: 0.14343059577230782 |
|
- ROUGE-l: 0.32809541498574013 |
|
|
|
### Framework versions |
|
|
|
- Transformers 4.27.1 |
|
- Pytorch 2.0.1+cu118 |
|
- Datasets 2.9.0 |
|
- Tokenizers 0.13.3 |
