license: apache-2.0
language: en
tags:
- generated_from_keras_callback
model-index:
- name: pull_request_comments_model
results: []
widget:
- text: Please add a detailed comment explaining what this config parameter means.
example_title: Code example
- text: >-
Because if non-chief workers don’t restore the model from the checkpoint,
training can‘t continue. For example, I am using tf.estimator.Estimator
API and tf.distribute.CollectiveAllReduceStrategy for distributed
training. Firstly, I train for 1000 rounds(train_stpes=1000) and save the
checkpoint, it works normally. Then I set the train_steps to 2000, only
the is_chief role can restore the model from the checkpoint without any
error.
example_title: ML example
- text: >-
Hi, Could you please review this PR when you have some time? Thank you
very much!
example_title: Management example
- text: >-
Sorry, i was on holiday all the month and i didn’t have so much time to
review this PR. I also had a lot of problems with my computer, so i
couldn’t work...
example_title: Other example
pipeline_tag: text-classification
pull_request_comments_model
Model description
This model is a fine-tuned version of distilbert-base-uncased on a pull request comments dataset. It achieves the following results on the evaluation set:
- Train Loss: 0.0791
- Train Accuracy: 0.9955
- Validation Loss: 0.5019
- Validation Accuracy: 0.8291
- Epoch: 12
Training and evaluation data
Training and evaluation data used for this model are the pull request comments of the tensorflow repository on GitHub. In particular, of all the pull request data (commit comments, review comments, events, exc.) only the rows with Type equal to PC (Pull request Comment) or RC (Review Comment) have been entered into the dataset. These comments has been classified into 4 categories:
- ML (Machine Learning), if the comment is about specific machine learning aspects, algorithms exc.
- Code, if the comment concerns either style and documentation in the code or maintainability issues or possible bugs exc.
- Management, if the comment is about management activities like checking an activity status, assign a review to someone, trigger Jenkins CI
- Other, if the comment doesn't belong to any of the above categories
Intended uses & limitations
One possible use of this model could be to label the pull request comments, clearly only on GitHub repositories that are about Machine Learning. In this way a developer, before reading a comment entirely, can have a preview of what that comment is about.
Training hyperparameters
The following hyperparameters were used during training:
- batch_size: 32
- num_epochs: 20
- optimizer: {'name': 'Adam', 'learning_rate': {'class_name': 'PolynomialDecay', 'config': {'initial_learning_rate': 2e-05, 'decay_steps': 280, 'end_learning_rate': 0.0, 'power': 1.0, 'cycle': False, 'name': None}}, 'decay': 0.0, 'beta_1': 0.9, 'beta_2': 0.999, 'epsilon': 1e-08, 'amsgrad': False}
- training_precision: float32
Training results
| Train Loss | Train Accuracy | Validation Loss | Validation Accuracy | Epoch |
|---|---|---|---|---|
| 1.3102 | 0.4308 | 1.1851 | 0.4701 | 0 |
| 1.1436 | 0.4978 | 0.9891 | 0.6068 | 1 |
| 0.9590 | 0.6183 | 0.8287 | 0.6838 | 2 |
| 0.7801 | 0.6942 | 0.6916 | 0.7692 | 3 |
| 0.6074 | 0.7946 | 0.6212 | 0.8120 | 4 |
| 0.4755 | 0.8817 | 0.5471 | 0.8205 | 5 |
| 0.3503 | 0.9241 | 0.5244 | 0.8376 | 6 |
| 0.2594 | 0.9665 | 0.5171 | 0.8120 | 7 |
| 0.1711 | 0.9844 | 0.4832 | 0.8291 | 8 |
| 0.1474 | 0.9911 | 0.5000 | 0.8205 | 9 |
| 0.1082 | 0.9955 | 0.4875 | 0.8291 | 10 |
| 0.0981 | 0.9933 | 0.4928 | 0.8291 | 11 |
| 0.0791 | 0.9955 | 0.5019 | 0.8291 | 12 |
Framework versions
- Transformers 4.26.1
- TensorFlow 2.9.0
- Datasets 2.10.1
- Tokenizers 0.13.2