|
--- |
|
base_model: google/flan-t5-large |
|
library_name: peft |
|
datasets: |
|
- jhu-clsp/jfleg |
|
language: |
|
- en |
|
pipeline_tag: text2text-generation |
|
tags: |
|
- text-generation-inference |
|
- grammar |
|
--- |
|
|
|
This model is part of the [GrammarCorrector](https://github.com/akhmat-s/GrammarCorrector) tool. |
|
|
|
"[FlanT5 from scratch for the grammar correction tool](https://medium.com/@akhmat-s/flant5-from-scratch-for-the-grammar-correction-tool-deadba9a6778)" article about how this models was trained: |
|
>FlanT5 was trained using [JFLEG](https://arxiv.org/abs/1702.04066) dataset. The primary objective of the experiment was to develop a highly effective tool using relatively small models, minimal datasets, and constrained computational resources. |
|
> |
|
>To accomplish this goal, we implemented two key strategies: |
|
>- [Perplexity-Based Data](https://arxiv.org/abs/2405.20541) Pruning With Small Reference Models. |
|
>- A simple sampling and voting method for [multiple LLM agents](https://arxiv.org/abs/2402.05120). model was trained. |
