search

sabithsn / APPDIA-Discourse-Style-Transfer

Data and code for APPDIA: A Discourse-aware Transformer-based Style Transfer Model for Offensive Social Media Conversations (COLING 2022).
12 stars 1 forks source link

Checkpoints for trained model and Evaluation script? #1

Open lihkinVerma opened 1 year ago

lihkinVerma commented 1 year ago

I was looking for the model checkpoint used for safe-score calculation, can you share that?

The script used for evaluating the responses generated for non-offensive text generation are missing in the repo, Can you please share that too?

Can you please share the code for training BART and T5?

sabithsn commented 1 year ago

Hi,

Sorry about the late reply. I have been traveling.

  1. Unfortunately, we do not have local/global tags. we used that only for analysis purposes.
  2. We are not realeasing the model itself due to the complications of hosting models. But here are the hyperparameters. Model: bert-base-uncased, maxlen: 100, batch size:8, lr = 8e-5, epochs: 3. we use the official splits of the OLID data. We used ktrain library for training. following these steps should yield a classifier that is very similar.
  3. Once more, we are not releasing the models themselves. Code for training the discourse-aware/baseline versions are in the repo. However, I am sharing our model output and evaluation script. I think the current bertscore model is a bit different from the one used 1 year ago. But the numbers should be in the same ballpark and you should be able to see very similar differences as reported in the paper.

Best, Sabit

On Wed, May 10, 2023 at 12:46 AM Nikhil Verma @.***> wrote:

I was looking for the model checkpoint used for safe-score calculation, can you share that?

The script used for evaluating the responses generated for non-offensive text generation are missing in the repo, Can you please share that too?

Can you please share the code for training BART and T5?

— Reply to this email directly, view it on GitHub https://github.com/sabithsn/APPDIA-Discourse-Style-Transfer/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANNRZK2DSD2ZIVLAYTDJ22DXFKGHNANCNFSM6AAAAAAX3WFGBI . You are receiving this because you are subscribed to this thread.Message ID: @.***>