llm-to-corpus 0.0.3

Large language model to corpus

Introduction

The goal of this tool is to apply Large Language Models operations to monolingual corpus to generate parallell corpus.

Uses cases:

• Asking a model to translate, summarize, paraphrasing original sentence to be able to benchmark its performance
• For corpus generation tasks from monolingual corpus, like for example, translated corpus.
• When developing prompts for your application, enables to test the prompt over a list of sentence to do evaluations

You basically provide an input file and prompt and it generates a target corpus:

Quick start

For example, to use OpenAI ChatGPT to translate a file:

llm-to-corpus samples/eng.txt samples/fra.txt "translate to French"

To see models and options available:

llm-to-corpus --help

Usage

Evaluation with Chatgpt

Translate Flores200 corpus to evalute quality of Catalan translation

llm-to-corpus samples/flores200.eng chatgpt.txt "Translate to Catalan the following text:"

pip install sacrebleu

sacrebleu samples/flores200.cat -i chatgpt.txt -m bleu chrf --format text

Evaluation with Bloom

Translate Flores200 corpus to evalute quality of Catalan translation

llm-to-corpus samples/flores200.eng bloom.txt "Translate to Catalan the following text:" --model mt0-xxl-mt

pip install sacrebleu

sacrebleu samples/flores200.cat -i bloom.txt -m bleu chrf --format text