/precise-outline-gen

Primary LanguagePython

Precise outline-conditioned text generation

Implementation for our work: Advancing Precise Outline-Conditioned Text Generation with Task Duality and Explicit Outline Control. (EACL24)

Environment

  • Python 3.9
  • datasets 1.8.0
  • transformers 4.33.2
  • torch 2.0.0
  • tokenizers 0.13.3
  • numpy 2.4.0

Usage

Data

Datasets available in Google Drive.

Code

Here we provide code for two scenarios. Fine-tuned methods are in the folderft_lm_conditional_gen. Please change the parameters(e.g., path) in outline_based_generation.sh. Run it directly by following script:

python outline_whole_generation.py 
    --model_name_or_path facebook/bart-base 
    --do_train 
    --do_predict 
    --train_file= ./data/WPOG/train.csv 
    --validation_file= .data/WPOG/validation.csv 
    --test_file= .data/WPOG/test.csv 
    --num_train_epochs=10 
    --max_source_length=512
    --max_target_length=1024
    --output_dir ./result/ 
    --per_device_train_batch_size=8 
    --per_device_eval_batch_size=8 
    --save_steps=30000 
    --predict_with_generate 
    --overwrite_output_dir

The zero-shot inference(LLMs) could be found at llm_inference.py. You could directly run by

python llm_inference.py