---
license: mit
library_name: transformers
tags:
- code
---

## JonBERTa-attn-ft-coco-234L

Model for the paper [**"A Transformer-Based Approach for Smart Invocation of Automatic Code Completion"**](https://arxiv.org/abs/2405.14753). 

#### Description
This model is fine-tuned on a code-completion dataset collected from the open-source [Code4Me](https://github.com/code4me-me/code4me) plugin. The training objective is to have a small, lightweight transformer model to filter out unnecessary and unhelpful code completions. To this end, we leverage the in-IDE telemetry data, and integrate it with the textual code data in the transformer's attention module. 

- **Developed by:** [AISE Lab](https://www.linkedin.com/company/aise-tudelft/) @ [SERG](https://se.ewi.tudelft.nl/), Delft University of Technology 
- **Model type:** [JonBERTa](https://github.com/Ar4l/curating-code-completions/blob/main/modeling_jonberta.py)
- **Language:** Code 
- **Finetuned from model:** [`CodeBERTa-small-v1`](https://huggingface.co/huggingface/CodeBERTa-small-v1). 

Models are named as follows: 

- `CodeBERTa` → `CodeBERTa-ft-coco-[1,2,5]e-05lr`
    - e.g. `CodeBERTa-ft-coco-2e-05lr`, which was trained with learning rate of `2e-05`.
- `JonBERTa-head` → `JonBERTa-head-ft-[dense,proj,reinit]` 
    - e.g. `JonBERTa-head-ft-dense-proj`, where all have `2e-05` learning rate, but may differ in the head layer in which the telemetry features are introduced (either `head` or `proj`, with optional `reinit`ialisation of all its weights).
- `JonBERTa-attn` → `JonBERTa-attn-ft-[0,1,2,3,4,5]L`
    - e.g. `JonBERTa-attn-ft-012L` , where all have `2e-05` learning rate, but may differ in the attention layer(s) in which the telemetry features are introduced (either `0`, `1`, `2`, `3`, `4`, or `5L`).

Other hyperparameters may be found in the paper or the replication package (see below).

#### Sources 

- **Replication Repository:** [`Ar4l/curating-code-completions`](https://github.com/Ar4l/curating-code-completions/tree/main)
- **Paper:** [**"A Transformer-Based Approach for Smart Invocation of Automatic Code Completion"**](https://arxiv.org/abs/2405.14753) 
- **Contact:** https://huggingface.co/Ar4l

To cite, please use 

```bibtex
@misc{de_moor_smart_invocation_2024,
	title = {A {Transformer}-{Based} {Approach} for {Smart} {Invocation} of {Automatic} {Code} {Completion}},
	url = {http://arxiv.org/abs/2405.14753},
	doi = {10.1145/3664646.3664760},
	author = {de Moor, Aral and van Deursen, Arie and Izadi, Maliheh},
	month = may,
	year = {2024},
}
```

#### Training Details 
This model was trained with the following hyperparameters, everything else being `TrainingArguments`' default. The dataset was prepared identically across all models as detailed in the paper. 

```python
num_train_epochs : int      = 3 
learning_rate    : float    = 2e-5
batch_size       : int      = 16
```

#### Model Configuration

```python
num_telemetry_features  :int = 26

add_feature_embeddings :bool = True 
feature_hidden_size     :int = num_telemetry_features * 4
feature_dropout_prob  :float = 0.1
add_feature_bias       :bool = True

add_self_attn          :bool = True
self_attn_layers  :list[int] = search(sum(
    [[i,j,k] for i in range(6) for j in range(6) for k in range(6) if i < j < k], 
    [[i,j] for j in range(6) for i in range(6) if i < j],
    [[i] for i in range(6)],
    []
))
```