h2oai/db_dirs

h2oGPT DataBase Data Card

Summary

H2O.ai's Chroma database files for h2oGPT for LangChain integration. Sources are generated and processed by get_db()

File	Purpose	Source	License
db_dir_DriverlessAI_docs.zip	DriverlessAI Documentation Q/A	Source	CC-BY-NC
db_dir_UserData.zip	Example PDFs and Text Files Q/A	Source	ArXiv
db_dir_github_h2oGPT.zip	h2oGPT GitHub repo Q/A	Source	Apache V2
db_dir_wiki.zip	Example subset of Wikipedia (from API) Q/A	Source	Wikipedia CC-BY-SA
db_dir_wiki_full.zip	All Wikipedia as of 04/01/2023 for articles with >5k views for Q/A	Source	Wikipedia CC-BY-SA

UserData can be generated for any collection of private offline docs by running make_db.py. For quickly using a private document collection for Q/A, place documents (PDFs, text, etc.) into a folder called user_path and run

python make_db.py

To use the chatbot with such docs, run:

python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6.9b --langchain_mode=UserData

using h2oGPT . Any other instruct-tuned base model can be used, including non-h2oGPT ones, as long as required GPU memory is avaialble for given model size. Or one can choose 8-bit generation.

See also LangChain example use with test_langchain_simple.py

If one has obtained all databases (except wiki_full) and unzipped them into the current directory, then one can run h2oGPT Chatbot like:

python generate.py --base_model=h2oai/h2ogpt-oasst1-512-12b --load_8bit=True --langchain_mode=UserData --visible_langchain_modes="['UserData', 'wiki', 'MyData', 'github h2oGPT', 'DriverlessAI docs']"

which uses now 12B model in 8-bit mode, that fits onto single 24GB GPU.

If one has obtained all databases (including wiki_full) and unzipped them into the current directory, then one can run h2oGPT Chatbot like:

python generate.py --base_model=h2oai/h2ogpt-oasst1-512-12b --load_8bit=True --langchain_mode=wiki_full --visible_langchain_modes="['UserData', 'wiki_full', 'MyData', 'github h2oGPT', 'DriverlessAI docs']"

which will default to wiki_full for QA against full Wikipedia.