pretty_name: 'JBCS2025: AES Experimental Logs and Predictions'
license: cc-by-nc-4.0
configs:
- config_name: evaluation_results
data_files:
- split: evaluation_results
path: evaluation_results-*.parquet
- config_name: bootstrap_confidence_intervals
data_files:
- split: boostrap_confidence_intervals
path: boostrap_confidence_intervals-*.parquet
tags:
- automatic-essay-scoring
- portuguese
- text-classification
JBCS 2025: Experimental Artefacts for AES in Brazilian Portuguese
This repository contains all experimental artefacts (logs, configurations, predictions, and evaluation results) described in the paper:
Exploring the Usage of LLMs for Automatic Essay Scoring in Brazilian Portuguese Essays
André Barbosa, Igor Cataneo Silveira, Denis Deratani Mauá
TODO
📦 What's in this dataset repo?
This dataset is not a training dataset. Instead, it provides comprehensive logs and outputs from experiments evaluating different language models for Automatic Essay Scoring (AES) tasks in Brazilian Portuguese.
Specifically, it contains:
- 🔁 JSONL files: raw predictions from each evaluated model.
- 📊 CSV files: detailed performance metrics (Quadratic Weighted Kappa, F1-score, etc.).
- ⚙️ YAML files: complete Hydra configurations for reproducibility.
- 📋 Log files: logs detailing each evaluation run.
📚 Related Collection
All models and datasets related to this work are available in the Hugging Face collection:
📊 Evaluated Models
The table below lists all models trained and evaluated for each essay competence (C1 to C5), along with direct links to their Hugging Face repository pages:
| Model | Architecture | Training Type | Link |
|---|---|---|---|
| mbert_base-C1 | Encoder-only | Fine-tuned | mbert_base-C1 |
| mbert_base-C2 | Encoder-only | Fine-tuned | mbert_base-C2 |
| mbert_base-C3 | Encoder-only | Fine-tuned | mbert_base-C3 |
| mbert_base-C4 | Encoder-only | Fine-tuned | mbert_base-C4 |
| mbert_base-C5 | Encoder-only | Fine-tuned | mbert_base-C5 |
| bertimbau_base-C1 | Encoder-only | Fine-tuned | bertimbau_base-C1 |
| bertimbau_base-C2 | Encoder-only | Fine-tuned | bertimbau_base-C2 |
| bertimbau_base-C3 | Encoder-only | Fine-tuned | bertimbau_base-C3 |
| bertimbau_base-C4 | Encoder-only | Fine-tuned | bertimbau_base-C4 |
| bertimbau_base-C5 | Encoder-only | Fine-tuned | bertimbau_base-C5 |
| bertimbau_large-C1 | Encoder-only | Fine-tuned | bertimbau_large-C1 |
| bertimbau_large-C2 | Encoder-only | Fine-tuned | bertimbau_large-C2 |
| bertimbau_large-C3 | Encoder-only | Fine-tuned | bertimbau_large-C3 |
| bertimbau_large-C4 | Encoder-only | Fine-tuned | bertimbau_large-C4 |
| bertimbau_large-C5 | Encoder-only | Fine-tuned | bertimbau_large-C5 |
| llama3-8b-C1 | Decoder-only | LoRA | llama3-8b-C1 |
| llama3-8b-C2 | Decoder-only | LoRA | llama3-8b-C2 |
| llama3-8b-C3 | Decoder-only | LoRA | llama3-8b-C3 |
| llama3-8b-C4 | Decoder-only | LoRA | llama3-8b-C4 |
| llama3-8b-C5 | Decoder-only | LoRA | llama3-8b-C5 |
| phi3.5-C1 | Decoder-only | LoRA | phi3.5-C1 |
| phi3.5-C2 | Decoder-only | LoRA | phi3.5-C2 |
| phi3.5-C3 | Decoder-only | LoRA | phi3.5-C3 |
| phi3.5-C4 | Decoder-only | LoRA | phi3.5-C4 |
| phi3.5-C5 | Decoder-only | LoRA | phi3.5-C5 |
| phi4-C1 | Decoder-only | LoRA | phi4-C1 |
| phi4-C2 | Decoder-only | LoRA | phi4-C2 |
| phi4-C3 | Decoder-only | LoRA | phi4-C3 |
| phi4-C4 | Decoder-only | LoRA | phi4-C4 |
| phi4-C5 | Decoder-only | LoRA | phi4-C5 |
🧠 Additionally, API-only models (e.g., DeepSeek-R1, ChatGPT-4o, Sabiá-3) were evaluated but are not hosted on the Hub. Their predictions and logs are still included in this dataset.
🧪 How to Use this Dataset
You can easily load the data using Hugging Face datasets library:
from datasets import load_dataset
ds = load_dataset("kamel-usp/jbcs2025_experiments", split="runs")
📄 License and Citation
This work is licensed under the Creative Commons Attribution 4.0 International License (CC-BY-4.0).
If you use these artefacts, please cite our paper:
TODO