This is the checkpoint for
RACE model, fine-tuned for the commit message generation (and/or completion) task as part of the paper "From Commit Message Generation to History-Aware Commit Message Completion", ASE 2023.
-
This model is based on the fine-tuned CodeT5 checkpoint
JetBrains-Research/cmg-codet5-without-history and uses RACE architecture introduced in 📜
RACE: Retrieval-Augmented Commit Message Generation.
- Note: Requires a custom model class. Check our implementation or the replication package provided by RACE authors.
-
This model was trained with commit diffs, WITHOUT commit message history.
-
This model was trained on the CommitChronicle dataset introduced in our study.
-
Our hyperparameter setting is mostly based on 📜
RACE: Retrieval-augmented Commit Message Generation.
The exact values are provided below:
| Hyperparameter | Value |
|---|
| Encoder context max length | 512 |
| Decoder context max length | 512 |
| Number of training epochs | 1 |
| Batch size | 32 |
| Optimizer | AdamW |
| Warmup | Linear |
| Number of warmup steps | 100 |
| Peak learning rate | 0.00002 |
We also released checkpoints for other models fine-tuned as part of our study.
