Introduction:
This research project explores LLM behavior on code-switched corpora in English and Spanish. Foundational work in this field includes the LinCE benchmark on code-switching tasks. Much of the existing data in this field is pulled from short social media interactions on Twitter (now X). Furthermore, LinCE does not explore generative tasks. Due to the rise in generative models, examining generative LLM behavior on similar data is important. Finally, human speech data is available. Short tweets or posts may not accurately represent human interactions. Using the Bangor Miami Corpus, I can validate and/or fine-tune Sagor Sarker's code-switched models (github: sagorbrur). In this way, I can expand on existing research towards both generative tasks and human data.
Files: job*.slurm files represent requests for compute on my clusters. You can use these files as guidelines for RAM, CPU, GPU needs for the code I am running.
prompting.py files are different kinds of prompts to the Mistral-8x7b Intruct model. They can be with Spanish, English, or a Mix.
The databricks files represent massive instruction datasets for LLMs. In addition to the human speech data, this is additional data that this project may use.
(https://huggingface.co/datasets/databricks/databricks-dolly-15k).
prompts.tsv pulls prompts from a similar study on South East Asian languages. Any references to languages that are out of scope (not English/Spanish) are modified to refer to the English/Spanish language pair. Any cultural and geographic references are similarly modified.
References: Sagor Sarker: https://github.com/sagorbrur/codeswitch https://huggingface.co/sagorsarker/codeswitch-spaeng-lid-lince/tree/main
Aguilar, Kar, Solorio: https://arxiv.org/abs/2005.04322
Bangor Miami Corpus: http://bangortalk.org.uk/speakers.php?c=miami
Prompting Data: https://github.com/Southeast-Asia-NLP/LLM-Code-Mixing