Wals Roberta Sets 136zip New Guide

Loop over the 136 test sets and aggregate metrics.

This could be a data release or a model checkpoint where:

In our internal testing, the 136zip set showed a 15-20% improvement in inference time compared to the previous 128 build, while maintaining comparable accuracy on standard GLUE benchmarks. wals roberta sets 136zip new

If "136zip" is an archive for a RoBERTa-related release, expected files:

If "sets" implies multiple parameter/config sets, the archive may include subfolders like: Loop over the 136 test sets and aggregate metrics

import json
from datasets import Dataset
def load_wals_roberta_set(path):
with open(path) as f:
data = json.load(f)
# assuming keys: 'input_ids', 'attention_mask', 'labels'
return Dataset.from_dict(data)

from transformers import RobertaForSequenceClassification, Trainer
model = RobertaForSequenceClassification.from_pretrained("roberta-base", num_labels=<num_features>)
trainer = Trainer(model=model, train_dataset=train_set, eval_dataset=dev_set)
trainer.train()

[Link to wals_roberta_sets_136zip.zip (2.3 GB)]