Dataset Construction

1. The images are under dataset_construction/dataset_images The code for generating the images is in dataset_construction/CreateImage.py Note: You might need to install a variety of fonts in the system in order to run the code.

2. Run generate_qa.ipynb to generate the full dataset and save them to a local dataset.

Running Experiments

Our code for running experiments rely on prl_ml(https://github.com/nuprl/prl_ml), which supports resumption.

Keep batch-size as 1 for all experiments, as batching is not possible for those models.

To run the experiments, first clone the prl_ml repository. Then drag all the files in run_model directory into prl_ml/batched_lm_generation directory. cd into the repository to run the generations with the commands below.

Multiple Choice

For each model, run the code file named by the model name. argument --fewshot-prompt: choose from '1shot', '3shot', '5shot', None(leave empty for zero-shot)

1. GPT-4o eg. Run GPT-4o using gpt4o_vision.py few-shot with three examples.

   python3 -m prl_ml.batched_lm_generation.gpt4o_vision \
       --dataset "disk:path_to_GlyphPattern_repo/dataset_construction/GlyphPattern/four_choice" \
       --output-dir gpt4o-4choice-zeroshot \
       --model-name gpt-4o \
       --temperature 0 \
       --batch-size 1 \
       --completion-limit 1 \
       --max-tokens 10 \
       --prompt-keys prompt,images \
       --extra-columns file_name,answer

2. Gemini 1.5 eg. Run Gemini1.5Pro using geminipro_vision.py few-shot with five examples.

   python3 -m batched_lm_generation.geminipro_vision \
       --dataset "disk:path_to_GlyphPattern_repo/dataset_construction/GlyphPattern/four_choice" \
       --output-dir geminipro-4choice-zeroshot \
       --model-name gemini-1.5-pro \
       --temperature 0 \
       --batch-size 1 \
       --completion-limit 1 \
       --max-tokens 10 \
       --prompt-keys prompt,images \
       --fewshot-prompt 5shot \
       --extra-columns file_name,answer

For Idefics models and other chat models that support AutoModelForVision2Seq Class, run the idefics_vision code file with output-dir and model-name replaced. 3. Idefics2 eg.Run Idefics2 from HuggingFaceM4/idefics2-8b few shot with one example:

```bash
python3 -m batched_lm_generation.idefics_vision \
    --dataset "disk:path_to_GlyphPattern_repo/dataset_construction/GlyphPattern/four_choice" \
    --output-dir Idefics2-4choice-zeroshot \
    --model-name HuggingFaceM4/idefics2-8b \
    --temperature 0 \
    --batch-size 1 \
    --completion-limit 1 \
    --prompt-keys prompt,images \
    --fewshot-prompt 1shot \
    --extra-columns file_name,answer
```

4. Idefics3 eg.Run Idefics3 from HuggingFaceM4/Idefics3-8B-Llama3 zero shot:

   python3 -m batched_lm_generation.idefics_vision \
       --dataset "disk:path_to_GlyphPattern_repo/dataset_construction/GlyphPattern/four_choice" \
       --output-dir Idefics2-4choice-zeroshot \
       --model-name HuggingFaceM4/idefics2-8b \
       --temperature 0 \
       --batch-size 1 \
       --completion-limit 1 \
       --prompt-keys prompt,images \
       --extra-columns file_name,answer

5. LlavaNext eg. Run LlavaNext from llava-hf/llava-v1.6-mistral-7b-hf Zero-shot

   python3 -m batched_lm_generation.llavanext_vision \
       --dataset "disk:path_to_GlyphPattern_repo/dataset_construction/GlyphPattern/four_choice" \
       --output-dir llavanext-4choice-zeroshot \
       --model-name llava-hf/llava-v1.6-mistral-7b-hf \
       --temperature 0 \
       --batch-size 1 \
       --completion-limit 1 \
       --prompt-keys prompt,images \
       --extra-columns file_name,answer

6. InstructBlip Run InstructBlip from Salesforce/instructblip-vicuna-7b, only zero-shot is supported.

   python3 -m batched_lm_generation.instructblip_vision \
       --dataset "disk:path_to_GlyphPattern_repo/dataset_construction/GlyphPattern/four_choice" \
       --output-dir instructblip-4choice-zeroshot \
       --model-name Salesforce/instructblip-vicuna-7b \
       --temperature 0 \
       --batch-size 1 \
       --completion-limit 1 \
       --prompt-keys prompt,images \
       --extra-columns file_name,answer

Few-shot Free Response

We support few-shot free response for the best two performing models, gpt-4o and gemini 1.5. Run code files with suffix freeresponse, and use dataset-split gt_descriptions.

1. GPT-4o

   python3 -m batched_lm_generation.gpt4o_freeresponse \
       --dataset "disk:path_to_GlyphPattern_repo/dataset_construction/GlyphPattern/gt_descriptions" \
       --output-dir gpt4o-free \
       --model-name gpt-4o \
       --temperature 0 \
       --batch-size 1 \
       --completion-limit 1 \
       --max-tokens 500 \
       --prompt-keys prompt,images \
       --extra-columns file_name,answer

2. Gemini1.5Pro

   python3 -m batched_lm_generation.geminipro_freeresponse \
       --dataset "disk:path_to_GlyphPattern_repo/dataset_construction/GlyphPattern/gt_descriptions" \
       --output-dir geminipro-free-leftright \
       --model-name gemini-1.5-pro \
       --temperature 0 \
       --batch-size 1 \
       --completion-limit 1 \
       --max-tokens 500 \
       --prompt-keys prompt,images \
       --extra-columns file_name,answer

Process Model Results:

The model results csv files are in model_results directory.

To process the generated results from running experiment above, run the check_answers.py in process_result directory. Replace the result_dir with the model output directory created by running batched_lm_generation.

For example, to check the outputs of running llavanext zero shot multiple choice: bash python process_results/check_answers.py \ --result_dir llavanext-4choice-zeroshot \ --model_name llavanext This produces a csv file at where the result directory was.

To calculate accuracy base on visual output, run outputAccuracy.py with the file_path replaced.