# Experiment 2 – Latin Square 1: CCT5 & COME on MCMD (Redesigned)

This repository contains the artifacts for **Latin Square 1 of Experiment 2**, which corresponds to the **redesigned and reimplemented experiment** evaluated on the **MCMD dataset** using the DNN-based commit message generation baselines **CCT5** and **COME**. This experiment was conducted under a more explicit and controlled evaluation protocol than the original study by Wu et al. (2025).

***

## Models

### CCT5
CCT5 is a code-change-oriented pre-trained model built on top of the **T5 architecture**, initialized from **CodeT5** weights and further pre-trained on **CodeChangeNet** (~40GB, 1.5M diff/commit pairs). Released at ESEC/FSE 2023.

- Architecture: Encoder-decoder Transformer (`T5-base` → `CodeT5` → `CCT5`)
- Pre-training corpus: CodeChangeNet (code diffs paired with commit messages)
- For MCMD: fine-tuned checkpoint from original CCT5 authors, trained on MCMD training set

### COME
COME (Commit Message Generation with Modification Embedding) is a hybrid DNN system built on top of CodeT5 with three core components:
- **Modification embedding**: converts code changes into numerical vectors capturing code evolution
- **Fine-tuned CodeT5**: generates candidate commit messages from the embedded representation
- **SVM-based decision algorithm**: selects between the generated and retrieved candidate messages

Released at ISSTA 2023. Does not include additional large-scale pre-training beyond CodeT5.

- For MCMD: language-specific checkpoints released by original COME authors (one per language)

***

## Dataset

**MCMD** – Multilingual Commit Message Dataset

| Property | Details |
|----------|---------|
| Languages | Java, C++, C#, Python, JavaScript |
| Repositories | Top 100 most-starred GitHub repos per language (500 total) |
| Total commits | ~1,094,115 |
| Date range | Up to January 1st, 2022 |
| Split | 80% train / 10% validation / 10% test |
| Authors | Liu et al. (2020) |

***

## Repository Structure

Each run folder corresponds to a **programming language** evaluated in this Latin Square. Unlike Experiment 1, this experiment follows a more controlled protocol with explicit random seed documentation and multiple evaluation runs where applicable.

```
experiment2_ls1/
├── run_java/
│   ├── checkpoint/          # CCT5 and COME checkpoints for MCMD (Java)
│   ├── predictions/         # Generated commit messages on MCMD Java test set
│   └── metrics/             # BLEU, METEOR, ROUGE-L, CIDEr scores
├── run_cpp/
│   ├── checkpoint/
│   ├── predictions/
│   └── metrics/
├── run_csharp/
│   ├── checkpoint/
│   ├── predictions/
│   └── metrics/
├── run_python/
│   ├── checkpoint/
│   ├── predictions/
│   └── metrics/
└── run_javascript/
    ├── checkpoint/
    ├── predictions/
    └── metrics/
```

### `checkpoint/`
Contains the model checkpoint files used for evaluation. For MCMD, these are the fine-tuned checkpoints released by the original authors of CCT5 and COME, trained on the MCMD training set for the corresponding language.

### `predictions/`
Contains the generated commit messages produced by each model on the MCMD test set for the corresponding language. Files are stored as `.txt` with one prediction per line, aligned to the reference messages in the test set.

### `metrics/`
Contains the evaluation metric scores computed by comparing the predictions against the MCMD test set reference messages. Each file records BLEU, METEOR, ROUGE-L, and CIDEr scores per model and language under the redesigned evaluation protocol.

***

## Evaluation Metrics

| Metric | Description |
|--------|-------------|
| **BLEU** | Bilingual Evaluation Understudy — measures n-gram precision between generated and reference messages |
| **METEOR** | Metric for Evaluation of Translation with Explicit Ordering — extends BLEU with recall, stemming, and synonym matching via WordNet |
| **ROUGE-L** | Recall-Oriented Understudy for Gisting Evaluation (LCS variant) — measures longest common subsequence overlap |
| **CIDEr** | Consensus-based Image Description Evaluation — TF-IDF-weighted n-gram similarity against reference messages |

### Reference Results (Original Paper – Wu et al., 2025)

| Language | Model | BLEU | METEOR | ROUGE-L | CIDEr |
|----------|-------|------|--------|---------|-------|
| Java | CCT5 | 17.19 | 14.95 | 26.08 | 1.06 |
| Java | COME | 27.17 | 23.36 | 34.59 | 1.90 |
| C++ | CCT5 | 15.65 | 14.11 | 24.15 | 0.90 |
| C++ | COME | 27.29 | 23.29 | 33.33 | 1.91 |
| C# | CCT5 | 12.06 | 11.05 | 18.92 | 0.61 |
| C# | COME | 20.80 | 17.72 | 27.01 | 1.25 |
| Python | CCT5 | 15.12 | 13.70 | 23.79 | 0.85 |
| Python | COME | 23.17 | 19.99 | 30.48 | 1.50 |
| JavaScript | CCT5 | 19.76 | 17.51 | 28.73 | 1.33 |
| JavaScript | COME | 26.91 | 23.02 | 34.44 | 1.92 |
| **Average** | **CCT5** | **15.96** | **14.26** | **24.33** | **0.95** |
| **Average** | **COME** | **25.07** | **21.48** | **31.97** | **1.70** |

These values serve as the baseline reference for comparison with the results produced under the redesigned protocol.

***

## Methodological Differences from Experiment 1

This experiment was redesigned to address the validity and reproducibility concerns identified during the Experiment 1 reproduction phase:

- **Explicit random seed documentation** for all runs
- **Controlled evaluation procedure** with clearly specified script versions
- **Full documentation of execution conditions** (hardware, software versions, environment)
- **Explicit treatment of validity threats** at each stage of the evaluation

***

## References

- Wu et al. (2025). *An Empirical Study on Commit Message Generation with Large Language Models via In-Context Learning.* arXiv:2502.18904.
- Lin et al. (2023). *CCT5: A Code-Change-Oriented Pre-Trained Model.* ESEC/FSE 2023.
- He et al. (2023). *COME: Commit Message Generation with Modification Embedding.* ISSTA 2023.
- Liu et al. (2020). *MCMD dataset.*
- Vegas & Elbaum (2023). *Pitfalls in Experiments with DNN4SE.* ESEC/FSE 2023.