Installing

• run install.sh (to install hardware dependent torch automatically)

Manual Installing

For GPU:

poetry install --extras "gpu"

For CPU:

poetry install --extras "cpu"

Data Generation

Prompt to manually generate ground truth from Qwen 2.5 VL 72b (from the chat UI):

""" Please extract all the informations that you find on the document in a json format. Note: The json keys must be in english (snake case) and the values in the document's original language. """

TODO:

• in poetry add all dependencies related to training under [train] extra
• make all empty null or ""
• add a mechanism to artificially add null elements on documents to train the model to not answer when element is not present (tricky)
• wait for llama cpu to release support to run qwen2 vl 2b