---
tags:
- pytorch
---

Checkpoints accompanying the LREC-COLING 2024 paper [Improved Neural Protoform Reconstruction via Reflex Prediction](https://arxiv.org/abs/2403.18769). For more detail, see our [repository](https://github.com/cmu-llab/reranked-reconstruction).

Directory structure: `dataset/task/architecture/run` where:

- dataset is one of `chinese_wikihan2022` (WikiHan), `chinese_wikihan2022_augmented` (WikiHan-aug), `baxter` (Hóu), `Nromance_ipa` (Rom-phon), or `Nromance_orto` (Rom-orth)
- task is either reflex prediction (`p2d`) or reconstruction (`d2p`)
- architecture is one of `GRU` (baseline in `p2d`, Meloni et al (2021)'s GRU in `d2p`), `JambuGRU` (Arora et al. (2023)'s GRU), `JambuTransformer` (Arora et al. (2023)'s Transformer), `Transformer` (Kim et al. (2023)'s Transformer), or `GRU-BS` (Meloni et al (2021)'s GRU with beam search)
- run is a unique identifier for a specific run