Skip to content

info.md

Latest commit

 

History

History
54 lines (36 loc) · 3.47 KB

info.md

File metadata and controls

54 lines (36 loc) · 3.47 KB

AC-LoRA

This folder contains the additional data required to reproduce our experiments.

Here is a summary of its content:

RepLiQA

This folder contains the relevant data to reproduce the experiments on the RepLiQA dataset.

  • db: contains the vector-database
  • test: contains our test split used to evaluate AC-LoRA
  • train: contains the train files (4 datapoints per question, as described in the paper) we used to finetune the LoRAs.
  • db_file: contains the files we used to build the vector-store.

Most included config point to data/RepLiQA/LoRAs/. We do not include our finetuned LoRAs here. Please either insert your finetuned LoRAs or change the path in the config file.

FlanV2

As mentioned in the paper, we use this Flanv2 to compare AC-LoRA with related work. Mainly LoraRetriever. To this end we used their setup and their LoRAs.

Before running these experiments, please first download the LoRAs. They should at the end be stored under data/Flan/LoRAs/. If one wants to change the location of these, one needs to change the corresponding config file before running the experiments.

The Flan folder additionally contains:

  • db: the vector base used by AC-LoRA to retrieve the relevant LoRAs. For space reasons it is splitted. ⚠️ Important: you need to reconstruct it into index.pkl (e.g., cat data/Flan/db/index_part_* > data/Flan/db/index.pkl) and index.faiss (e.g., cat data/Flan/db/index_flan_part_* > data/Flan/db/index.faiss)
  • test: a copy of the test file from LoraRetriever.

WikiArts

This folder already contains everything needed to reproduce the experiments. In case one wants to re-finetune the LoRAs one needs to download the images from WikiArts

The folder contains the following data:

  • db: the vector store for wikiarts, constructed on the generated prompts.
  • test: some example prompts
  • train: contains the metadata of the generated prompts used to construct the db and the id of the corresponding image.

The relevant config points to data/WikiArts/LoRAs/. We do not include our finetuned LoRAs here. Please either insert your finetuned LoRAs or change the path in the config file.

MMSci

This folder contains the following data we used to run the experiments on our variant of the MMSci dataset:

  • db: the vector store constructed using the files in train and the images from the benchmark/train as described from MMSci. These images are not required exept one wants to finetune the LoRAs or recompute the db.
  • train: contains our train split used to build the db.
  • test: contains the test file from benchmark\dev\ we used to evaluate the retrival capacities. To run the experiments on needs to follow the instructions from MMSci and download the images from benchmark\dev\.

CS-Combi

This folder contain the generated dataset we used to evaluate the capacity from AC-LoRA to combine information. This folder contains:

  • test: contains the test set as described in the paper.
  • train: contains the two train set to train the LoRAs1 and LoRAs2.