README.md

metadata

language:
  - ru
tags:
  - fluency

This is a ruRoberta-large model trained on the RuCoLa dataset. It can be used to classify Russian sentences into fluent or non-fluent ones, where fluency is understood as linguistic acceptability.

Training notebook: task_oriented_TST/fluency/rucola_classifier_v1.ipynb (in a private repo).

Training parameters:

• optimizer: Adam
• lr=2e-6
• batch_size=32
• epochs=10
• clip_grad_norm=1.0

Test accuracy (on the leaderboard this model is submitted as ruroberta-base-cased-rucola-v1): 0.81.