AKG2/flan-t5-base-xsum-lora

FLAN-T5-Base Fine-tuned on XSum with LoRA

This model is a fine-tuned version of google/flan-t5-base on the XSum dataset using LoRA (Low-Rank Adaptation) for parameter-efficient fine-tuning.

Model Description

• Base Model: google/flan-t5-base
• Task: Extreme Summarization (one-sentence summaries)
• Dataset: XSum (BBC news articles)
• Training Method: LoRA (Low-Rank Adaptation)
• Parameters: 0.00M trainable (0.00% of 249.35M total)

Training Details

LoRA Configuration

• Rank (r): 16
• Alpha: 32
• Target modules: q, v
• Dropout: 0.05

Training Hyperparameters

• Learning rate: 3e-4
• Batch size: 8
• Epochs: 3
• Optimizer: AdamW
• Mixed precision: FP16

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from peft import PeftModel

# Load base model and tokenizer
base_model = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-base")
tokenizer = AutoTokenizer.from_pretrained("AKG2/flan-t5-base-xsum-lora")

# Load LoRA adapters
model = PeftModel.from_pretrained(base_model, "AKG2/flan-t5-base-xsum-lora")

# Generate summary
text = "Your article text here..."
inputs = tokenizer("summarize: " + text, return_tensors="pt", max_length=512, truncation=True)
outputs = model.generate(**inputs, max_length=64, num_beams=4, length_penalty=2.0)
summary = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(summary)

Performance

Evaluation metrics on XSum test set:

• ROUGE-1: [Add your score]
• ROUGE-2: [Add your score]
• ROUGE-L: [Add your score]

Citation

If you use this model, please cite the original FLAN-T5 paper and the XSum dataset:

@article{chung2022scaling,
  title={Scaling instruction-finetuned language models},
  author={Chung, Hyung Won and others},
  journal={arXiv preprint arXiv:2210.11416},
  year={2022}
}

@inproceedings{narayan2018don,
  title={Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization},
  author={Narayan, Shashi and others},
  booktitle={EMNLP},
  year={2018}
}

License

This model inherits the license from the base model: Apache 2.0

---

Trained by: AKG2