hf_mirrors/roberta_large
Model Introduction
Files & versions
Analysis

Model description

RoBERTa is a transformers model pretrained on a large corpus of English data in a self-supervised fashion. This means it was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots of publicly available data) with an automatic process to generate inputs and labels from those texts.

More precisely, it was pretrained with the Masked language modeling (MLM) objective. Taking a sentence, the model randomly masks 15% of the words in the input then run the entire masked sentence through the model and has to predict the masked words. This is different from traditional recurrent neural networks (RNNs) that usually see the words one after the other, or from autoregressive models like GPT which internally mask the future tokens. It allows the model to learn a bidirectional representation of the sentence.

This way, the model learns an inner representation of the English language that can then be used to extract features useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a standard classifier using the features produced by the BERT model as inputs.

This bias will also affect all fine-tuned versions of this model.

Training data

The RoBERTa model was pretrained on the reunion of five datasets:

  • [BookCorpus], a dataset consisting of 11,038 unpublished books;
  • [English Wikipedia] (excluding lists, tables and headers) ;
  • [CC-News], a dataset containing 63 millions English news articles crawled between September 2016 and February 2019.
  • [OpenWebText], an opensource recreation of the WebText dataset used to train GPT-2,
  • [Stories] a dataset containing a subset of CommonCrawl data filtered to match the story-like style of Winograd schemas.

Together theses datasets weight 160GB of text.

Use with openMind

environment variable

# source environment variable
source /usr/local/Ascend/ascend-toolkit/set_env.sh
export OPENMIND_FRAMEWORK=pt

pip install openMind Library

OpenMind Library can be installed through pip, please select the corresponding command according to the actual environment for installation.

It should be noted that since torch npu depends on torch, which can be directly installed through pip in aarch64, but requires a specific URL to download the CPU version in x86, the installation commands in the two environments are different. The specific installation code has been distinguished and presented in the following text.

# aarch64
pip install openmind[all]
# x86
pip install openmind[all] --extra-index-url https://download.pytorch.org/whl/cpu

Inference

from openmind import AutoTokenizer, AutoModelForCausalLM
import torch
import torch_npu

model_dir = "HangZhou_Ascend/roberta_large"
tokenizer = AutoTokenizer.from_pretrained(model_dir, device_map="auto", trust_remote_code=True)
# Set `torch_dtype=torch.float16` to load model in float16, otherwise it will be loaded as float32 and might cause OOM Error.
model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="auto",  trust_remote_code=True, torch_dtype=torch.float16)
model = model.eval()
response, history = model.chat(tokenizer, "1+1=", history=[], meta_instruction="")
print(response)

Preprocessing

The texts are tokenized using a byte version of Byte-Pair Encoding (BPE) and a vocabulary size of 50,000. The inputs of the model take pieces of 512 contiguous token that may span over documents. The beginning of a new document is marked with <s> and the end of one by </s>

The details of the masking procedure for each sentence are the following:

  • 15% of the tokens are masked.

  • In 80% of the cases, the masked tokens are replaced by <mask>.

  • In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace.

  • In the 10% remaining cases, the masked tokens are left as is.

Contrary to BERT, the masking is done dynamically during pretraining (e.g., it changes at each epoch and is not fixed).

Pretraining

The model was trained on 1024 V100 GPUs for 500K steps with a batch size of 8K and a sequence length of 512. The optimizer used is Adam with a learning rate of 4e-4, $\beta_{1} = 0.9$, $\beta_{2} = 0.98$ and $\epsilon = 1e-6$, a weight decay of 0.01, learning rate warmup for 30,000 steps and linear decay of the learning rate after.

Evaluation results

When fine-tuned on downstream tasks, this model achieves the following results:

Glue test results:

TaskMNLIQQPQNLISST-2CoLASTS-BMRPCRTE
90.292.294.796.468.096.490.986.6