XLM RoBERTa Base ANCE-Warmuped

This is a XLM RoBERTa Base model trained with ANCE warmup script. RobertaForSequenceClassification is replaced to XLMRobertaForSequenceClassification in warmup script.

trained 60k steps.

train args is below:

data_dir: ../data/raw_data/
train_model_type: rdot_nll
model_name_or_path: xlm-roberta-base
task_name: msmarco
output_dir: 
config_name: 
tokenizer_name: 
cache_dir: 
max_seq_length: 128
do_train: True
do_eval: False
evaluate_during_training: True
do_lower_case: False
log_dir: ../logs/
eval_type: full
optimizer: lamb
scheduler: linear
per_gpu_train_batch_size: 32
per_gpu_eval_batch_size: 32
gradient_accumulation_steps: 1
learning_rate: 0.0002
weight_decay: 0.0
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 2.0
max_steps: -1
warmup_steps: 1000
logging_steps: 1000
logging_steps_per_eval: 20
save_steps: 30000
eval_all_checkpoints: False
no_cuda: False
overwrite_output_dir: True
overwrite_cache: False
seed: 42
fp16: True
fp16_opt_level: O1
expected_train_size: 35000000
load_optimizer_scheduler: False
local_rank: 0
server_ip: 
server_port: 
n_gpu: 1
device: cuda:0
output_mode: classification
num_labels: 2
train_batch_size: 32

Eval Result

Reranking/Full ranking mrr: 0.27380855732933/0.24284821712830248
{"learning_rate": 0.00019460324719871943, "loss": 0.0895877162806064, "step": 60000}

Usage

from transformers import XLMRobertaForSequenceClassification, XLMRobertaTokenizer
repo = "k-ush/xlm-roberta-base-ance-warmup"
model = XLMRobertaForSequenceClassification.from_pretrained(repo)
tokenizer = XLMRobertaTokenizer.from_pretrained(repo)