Tasks | Datasets | LongLM | Baselines | Paper

LOT is a benchmark for evaluating Chinese long text modeling. LOT consists of two understanding tasks and two generation tasks. We construct new datasets for these tasks based on human-written Chinese stories.

Furthermore, we release an encoder-decoder-based Chinese long text pretraining model named LongLM with up to 1 billion parameters. We pretrain LongLM on 120G Chinese novels with two generative tasks including text infilling and conditional continuation. Extensive experiments show that LongLM outperforms similar-sized pretraining models substantially on both the understanding and generation tasks in LOT.

We design LOT as an aggregation of two understanding tasks including Cloze Test (ClozeT) and Sentence Position Prediction (SenPos), and two generation tasks including Plot Completion (PlotCom) and Outline-conditioned Generation (OutGen). We show the task descriptions in the table below.

We show the data statistics in the table below. The abbreviation sent/len is short for sentence/length, respectively. The datasets and evaluation scripts can be downloaded from THUCloud.

• $d_m$: the dimension of hidden states
• $d_{ff}$: the dimension of feed forward layers
• $d_{kv}$: the dimension of the keys/values in the self-attention layers
• $n_h$: the number of attention heads
• $n_e$: the number of hidden layers of the encoder
• $n_d$: the number of hidden layers of the decoder
• #P: the number of parameters

We collect 120G novels as the pretraining data for LongLM. Part of the pretraining data are publicly available.

1. Download: The checkpoints and example data can be downloaded from THUCloud or Hugging Face Model Card. The training and generation scripts are under the directory longlm.

2. Model Loading:

   from transformers import T5Tokenizer, T5ForConditionalGeneration
   tokenizer = T5Tokenizer.from_pretrained('thu-coai/LongLM-large')
   model = T5ForConditionalGeneration.from_pretrained('thu-coai/LongLM-large')
   

3. Training:

   Execute bash ./finetune.sh to fine-tune LongLM. If deepspeed is available, you can execute bash ./finetune_deepspped.sh to accelerate. You can also use the official script provided by Transformers to fine-tune the model.

   env CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 CUDA_LAUNCH_BLOCKING=1 python3 -m torch.distributed.launch --nproc_per_node=8 \
   finetune_trainer.py \
   --data_dir=./data \ # directory of data
   --train_name=train \ # file prefix of the training data
   --output_dir=./save_model \ # output directory to save the checkpoint
   --save_total_limit=10 \ # maximum number of the saved checkpoints
   --per_gpu_train_batch_size=3 \ # batch size for training
   --per_gpu_eval_batch_size=3 \ # batch size for evaluation
   --num_train_epochs=1 \ # number of training epochs
   --logging_steps=5 \ # number of stps to log the loss value
   --model_name_or_path=./LongLM-small \ # path to the pretrained model
   --warmup_steps=100 \ # number of steps for warmup
   --learning_rate=1e-4 \ # learning rate
   --n_val=100 \ # number of examples for validation
   --do_train --do_eval \ # whether to training/validation
   --evaluation_strategy steps \ # strategy of evaluation
   --gradient_accumulation_steps=40 # number of steps for gradient accumulation
   --overwrite_output_dir \
   --load_best_model_at_end

4. Generation:

   input_ids = tokenizer("小咕噜对，<extra_id_1>",return_tensors="pt", padding=True, truncation=True, max_length=512).input_ids.to(device)
   
   gen = model.generate(input_ids, do_sample=True, decoder_start_token_id=1, top_p=0.9, max_length=512)

datasets                1.6.2
deepspeed               0.3.16
huggingface-hub         0.0.8
jieba                   0.42.1
jsonlines               2.0.0
nltk                    3.5
numpy                   1.19.5
pytorch-lightning       1.2.0
regex                   2020.11.13
rouge                   1.0.1
rouge-score             0.0.4
sacrebleu               1.5.0
scipy                   1.5.4
sentencepiece           0.1.95
tokenizers              0.10.1
torch                   1.8.1
torchaudio              0.8.0
torchmetrics            0.2.0
torchvision             0.9.0
transformers            4.6.1

The example data, training and evaluation scripts of LongLM are under the directory ./baselines/understanding. You can execute bash ./finetune.sh to fine-tune LongLM and execute bash ./eval.sh to evaluate the fine-tuned model.

The training script of LongLM for the generation tasks is the same as pretraining script. The generation script and example data can be found under ./baseline/generation. You can execute bash ./gen.sh for generation.

@misc{guan2021lot,
      title={LOT: A Benchmark for Evaluating Chinese Long Text Understanding and Generation}, 
      author={Jian Guan and Zhuoer Feng and Yamei Chen and Ruilin He and Xiaoxi Mao and Changjie Fan and Minlie Huang},
      year={2021},
      eprint={2108.12960},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}