polygloss

KO | EN

gitlite — search

#python #react #typescript #llm #go #android #rust #docker #cli #kubernetes #security #ios #mcp #open-source

(★ 15)

A massively multilingual corpus and pretrained model for IGT

polygloss 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# PolyGloss (formerly GlossLM)

A Massively Multilingual Corpus and Pretrained Model for Interlinear Glossed Text

- PolyGloss (v2): [📄 Paper](https://arxiv.org/abs/2601.10925) | [📦 Models and data](https://hf.co/collections/lecslab/polygloss)
- GlossLM (v1): [📄 Paper](https://arxiv.org/abs/2403.06399) | [📦 Models and data](https://huggingface.co/collections/lecslab/glosslm-66da150854209e910113dd87)

> [!NOTE]
> To see instructions for using the v1 GlossLM models, switch this repo to the `frozen/glosslm-v1` branch.

## Overview

PolyGloss consists of a multilingual corpus of IGT and pretrained models for interlinear glossing.

## Background

Interlinear Glossed Text (IGT) is a common format in language documentation projects. It looks something like this:

```
o sey x  -tok    r  -ixoqiil
o sea COM-buscar E3S-esposa
“O sea busca esposa.”
```

The three lines are as follows:

- The **transcription** line is a sentence in the target language (possibly segmented into morphemes)
- The **gloss** line gives an linguistic gloss for each morpheme in the transcription
- The **translation** line is a translation into a higher-resource language

Creating consistent and large IGT datasets is time-consuming and error-prone. The goal of **automated interlinear glossing** is to aid in this task by **predicting the gloss line given the transcription and translation line**.

## Using the dataset

```python
import datasets

polygloss_corpus = datasets.load_dataset("lecslab/polygloss-corpus")
```

## Using the interleaved model

```python
import transformers

model = transformers.AutoModelForSeq2SeqLM.from_pretrained("lecslab/polygloss-byt5-interleaved-2025-12-28", trust_remote_code=True)

# Estimate performance based on a glottocode
model.estimate_perf(glottocode="uspa1245")

# Predict IGT outputs
transcription = "o sey xtok rixoqiil"
translation = "O sea busca esposa."
lang = "Uspanteco"
metalang = "Spanish"

output = model.predict_igt(transcription, translation, lang, metalang)
print(outputs) # CONJ(o) sea(sey) COM(x)-buscar(tok) E3S(r)-esposa(ixoqiil)
```

## License

## Citation

PolyGloss:
```
@misc{ginn2026massivelymultilingualjointsegmentation,
      title={Massively Multilingual Joint Segmentation and Glossing}, 
      author={Michael Ginn and Lindia Tjuatja and Enora Rice and Ali Marashian and Maria Valentini and Jasmine Xu and Graham Neubig and Alexis Palmer},
      year={2026},
      eprint={2601.10925},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.10925}, 
}
```

GlossLM:
```
@inproceedings{ginn-etal-2024-glosslm,
    title = "{G}loss{LM}: A Massively Multilingual Corpus and Pretrained Model for Interlinear Glossed Text",
    author = "Ginn, Michael  and
      Tjuatja, Lindia  and
      He, Taiqi  and
      Rice, Enora  and
      Neubig, Graham  and
      Palmer, Alexis  and
      Levin, Lori",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.emnlp-main.683",
    pages = "12267--12286",
}
```

## Running Experiments

> [!NOTE]
> To reproduce the original GlossLM v1 experiments, please check out the `glosslm-v1` branch.

Interested in replicating or modifying our experiments? Pretraining, finetuning, and inference are handled with `run.py`, as follows:

```bash
# Python >=3.11
# Recommended to use venv for set up
python run.py some_config_file.cfg -o key1=val1 key2=val2
```

The run is defined by an INI-style _config file_. Each experiment should consist of a folder in `experiments` along with configuration files for pretraining, finetuning, or inference. See [an example experiment folder](experiments/base_igt). A config file looks something like:

```ini
[config]

mode = pretrain
pretrained_model = google/byt5-base

# Dataset
dataset_key = lecslab/polygloss-corpus

# Training
max_epochs = 50
learning_rate = 5e-5
batch_size = 16
```

The full list of possible options is in [experiment_config.py](src/training/experiment_config.py). In addition to the config file, you can specify any parameter overrides with `-o key1=val1 key2=val2`.

> [!NOTE]
> When running experiments over our evaluation languages, we use the paradigm shown in `run_curc.sh` where the config does not include a `glottocode` field and an override is used (e.g. `-o glottocode=uspa1245`).

## Development

### Set up environment
```bash
python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements
```

### Testing
Tests should be written as functions in modules prefixed with `test_`, as in `test_evaluate_segmentation_example` in `evaluate.py`. Run tests with:

```bash
PYTHONPATH=. pytest
```

### Dataset
The full dataset creation is handled by `python create_dataset.py`. Any changes to the dataset should be incorporated in this script or its child scripts, so we have a definitive source of truth.

polygloss

polygloss 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

polygloss

polygloss 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README