TinyLlama History Chat
Project Purpose
The purpose of this project is to provide a small language model fine-tuned with history information. To be easily deployed through an inference or local ui, and provides clear answers to historical questions.
Currently, the datasets that it has trained on are focused on American History.
Project's Current Progress
The current progress of the project is on v0.3. We have had other versions that jumped a head of that number, but the datasets used in those versions had some issues or major flaws that were cleaned up in later databases. This has brought us back to square one in the numbering.
- ambrosfitz/tinyllama-history-chat_v0.1
- This model was fine-tuned from the base Tinyllama v1.0 with a dataset of History v1. This dataset was comprised of questions and answers, generated by ChatGPT-3.5-Turbo, from the American Yawp Part 1, an open source US History Textbook.
- ambrosfitz/tinyllama-history-chat_v0.2
- This model was fine-tuned from the newly created Tinyllama-history-chat v0.1, using the History v2 dataset. This dataset comprises questions and answers from Chatgpt-3.5-Turbo, using the the American Yawp Part 2.
- ambrosfitz/tinyllama-history-chat_v0.3
- This model was fine-tuned from the newly created Tinyllama-history-chat v0.2, using the History v3 dataset. This dataset comprises questions and answers from Chatgpt-3.5-Turbo, using the OpenStax US History as it text data.
Limitations
As all language models, this is a trial by error in all cases. Some of the questions it provides are clearly incorrect, but other times it is perfectly correct. There are little safe guards on speech, so use at your own risk.
Next Steps
Currating datasets is the current path forward, finding and generating questions that have broad US History alignment. This takes time and money, which are both limited. We are currently using an API to run Q & A using Mixtral 8x7B model for most of these question. We plan to compare these results to ones against ChatGPT-3.5-Turbo and ChatGPT-4.