HuLU evaluate

hulu_evaluate is a library for evaluating and training language models on Hungarian tasks within the HuLU benchmark. It includes support for fine-tuning with LoRA, official evaluation scripts, and provides a leaderboard for benchmarking.

Features

• Training on multiple tasks: Supports tasks like "CoLA", "RTE", "WNLI", "CB", "SST", and "COPA".
• LoRA (Low-Rank Adaptation): Fine-tune models with reduced compute requirements.
• Official training scripts: Provided for training and evaluation.
• HuLU Leaderboard: Submit your results to the HuLU leaderboard.
• Fully Sharded Data Parallel: Shard and paralellize training of models

Installation

Installation via pypi

pip install hulu-evaluate

To install the hulu_evaluate library for testing and development, clone the repository and install dependencies:

pip install git+https://git.nlp.nytud.hu/DLT/HuLU-evaluate.git

Usage

Command-Line Interface (CLI) The CLI provides commands for configuration, training, and submission.

Setting Up the CLI The CLI entry point can be called with:

hulu-evaluate <command> [<args>]

Commands include

Train a Model: Train on HuLU tasks using the train command

hulu-evaluate train --model_name <MODEL_NAME> --output_dir <OUTPUT_DIR> --train_epochs 6 --train_batch 8

You can submit your results on the hulu.nytud.hu webpage. The results are created by default in the "HuluFinetune" directory.

Programmatic Usage

You can also integrate hulu_evaluate functionality directly in Python scripts.

Training

from hulu_evaluate.train.main import Main
from hulu_evaluate.hulu_arguments.train_arguments import Arguments

# Define training arguments
arguments = Arguments(
    model_name="my_model",
    train_epochs=6,
    tasks=["cola", "rte"],
)

# Train the model
Main(arguments)

Command Reference

Training Command

train - Trains a model on HuLU tasks with specified parameters.

hulu-evaluate train --model_name <MODEL_NAME> --output_dir <OUTPUT_DIR> --train_epochs <EPOCHS>
--model_name: Name of the model to train.
--train_epochs: Number of epochs.
--use_lora: Enable LoRA fine-tuning (optional).
--config_file: Path to config JSON file
--output_dir: Output directory
--model_name: Model name
--tokenizer_name: Tokenizer name (defaults to model_name)
--train_epochs:  default=6 Number of training epochs
--train_batch: default=8 Training batch size
--train_lr: default=2e-05 Learning rate
--train_warmup: default=0 Warmup steps
--train_maxlen: default=256 Max sequence length
--train_seed: default=42 Random seed
--precision: default="fp32" Precision (e.g., fp16 or fp32)
--use_lora: default=False Use LoRA for training
--lora_r: default=8 LoRA r parameter
--lora_alpha: default=16 LoRA alpha parameter
--lora_dropout: default=0.1 LoRA dropout rate
--tasks default=["cola", "rte", "wnli", "cb", "sst"] List of tasks to train on
--use_fsdp: default=False Using FSDP
--gradient_accumulation_steps: default=1 Set steps for gradient accumulation

The HuLU Benchmark

HuLU (Hungarian Language Understanding Benchmark Kit) was created on the basis of the GLUE and SuperGLUE benchmark databases. The main purpose of HuLU is to enable a standardized evaluation of neural language models in a simple way while also enabling multi-perspective analysis. It also compares the performance of various LMs on various tasks and shares the results with LT professionals in tabular and graphical formats. The tasks and their precise presentation is available on the official HuLU page

Citing

@inproceedings{hatvani2024hulu,
  author    = {Péter Hatvani and Kristóf Varga and Zijian Győző Yang},
  title     = {Evaluation Library for the Hungarian Language Understanding Benchmark (HuLU)},
  booktitle = {Proceedings of the 21th Hungarian Computational Linguistics Conference},
  year      = {2024},
  address   = {Hungary},
  publisher = {[Publisher Name]},
  note      = {Affiliations: PPKE Doctoral School of Linguistics, HUN-REN Hungarian Research Center for Linguistics},
  url       = {https://hatvanipeter.hu/},
  email     = {hatvani9823@gmail.com, varga.kristof@nytud.hun-ren.hu, yang.zijian.gyozo@nytud.hun-ren.hu}
}

Contributing

Contributions are welcome! Please submit issues or pull requests to improve hulu_evaluate.

License

This project is licensed under the Apache License.