philschmid/flan-t5-xxl-samsum-peft

FLAN-T5-XXL LoRA fine-tuned on samsum

PEFT tuned FLAN-T5 XXL model.

flan-t5-base-samsum

This model is a fine-tuned version of philschmid/flan-t5-xxl-sharded-fp16 on the samsum dataset. It achieves the following results on the evaluation set:

• rogue1: 50.386161%

• rouge2: 24.842412%

• rougeL: 41.370130%

• rougeLsum: 41.394230%

How to use the model

The model was trained using 🤗 PEFT. This repository only contains the fine-tuned adapter weights for LoRA and the configuration to load the model. Below you can find a snippet on how to run inference using the model. This will load the FLAN-T5-XXL from hugging face if not existing locally.

1. load the model

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

# Load peft config for pre-trained checkpoint etc. 
peft_model_id = "philschmid/flan-t5-xxl-samsum-peft"
config = PeftConfig.from_pretrained(peft_model_id)

# load base LLM model and tokenizer
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. generate

text = "test"

input_ids = tokenizer(text, return_tensors="pt", truncation=True).input_ids.cuda()
outputs = model.generate(input_ids=input_ids, max_new_tokens=10, do_sample=True, top_p=0.9)
print(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: 1e-3
• train_batch_size: auto
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 5

Framework versions

• Transformers 4.27.1
• Pytorch 1.13.1+cu117
• Datasets 2.9.1
• PEFT@main