[Question] Unable to reproduce GQA, SQA, and MME results on LLaVA-v1.5-7b

#!/bin/bash

gpu_list="${CUDA_VISIBLE_DEVICES:-0}"
IFS=',' read -ra GPULIST <<< "$gpu_list"

CHUNKS=${#GPULIST[@]}

CKPT="llava-v1.5-7b"
SPLIT="llava_gqa_testdev_balanced"
GQADIR="./playground/data/eval/gqa/data"

for IDX in $(seq 0 $((CHUNKS-1))); do
    CUDA_VISIBLE_DEVICES=${GPULIST[$IDX]} python -m llava.eval.model_vqa_loader \
        --model-path liuhaotian/llava-v1.5-7b \
        --question-file ./playground/data/eval/gqa/$SPLIT.jsonl \
        --image-folder ./playground/data/eval/gqa/data/images \
        --answers-file ./playground/data/eval/gqa/answers/$SPLIT/$CKPT/${CHUNKS}_${IDX}.jsonl \
        --num-chunks $CHUNKS \
        --chunk-idx $IDX \
        --temperature 0 \
        --conv-mode vicuna_v1 &
done

wait

output_file=./playground/data/eval/gqa/answers/$SPLIT/$CKPT/merge.jsonl

# Clear out the output file if it exists.
> "$output_file"

# Loop through the indices and concatenate each file.
for IDX in $(seq 0 $((CHUNKS-1))); do
    cat ./playground/data/eval/gqa/answers/$SPLIT/$CKPT/${CHUNKS}_${IDX}.jsonl >> "$output_file"
done

python scripts/convert_gqa_for_eval.py --src $output_file --dst $GQADIR/testdev_balanced_predictions.json

cd $GQADIR
python eval/eval.py --tier testdev_balanced

Binary: 79.11%                                 
Open: 47.23%                                   
Accuracy: 61.86%
Validity: 0.00%                                
Plausibility: 0.00%                            
Distribution: 1.64 (lower is better)
                                               
Accuracy / structural type: 
  choose: 83.35% (1129 questions)
  compare: 64.86% (589 questions)
  logical: 75.93% (1803 questions)
  query: 47.23% (6805 questions)
  verify: 83.26% (2252 questions)
                                               
Accuracy / semantic type:   
  attr: 68.20% (5186 questions)
  cat: 53.09% (1149 questions)
  global: 63.06% (157 questions)
  obj: 87.66% (778 questions)
  rel: 53.75% (5308 questions)
                                               
Accuracy / steps number:   
  1: 80.59% (237 questions)
  2: 56.54% (6395 questions)
  3: 64.96% (4266 questions)
  4: 68.47% (793 questions)
  5: 73.24% (822 questions)
  6: 82.93% (41 questions)                                                                    
  7: 100.00% (20 questions)
  8: 100.00% (3 questions)
  9: 100.00% (1 questions)

#!/bin/bash

python -m llava.eval.model_vqa_science \
    --model-path liuhaotian/llava-v1.5-7b \
    --question-file ./playground/data/eval/scienceqa/llava_test_CQM-A.json \
    --image-folder ./playground/data/eval/scienceqa/images/test \
    --answers-file ./playground/data/eval/scienceqa/answers/llava-v1.5-7b.jsonl \
    --single-pred-prompt \
    --temperature 0 \
    --conv-mode vicuna_v1

python llava/eval/eval_science_qa.py \
    --base-dir ./playground/data/eval/scienceqa \
    --result-file ./playground/data/eval/scienceqa/answers/llava-v1.5-7b.jsonl \
    --output-file ./playground/data/eval/scienceqa/answers/llava-v1.5-7b_output.jsonl \
    --output-result ./playground/data/eval/scienceqa/answers/llava-v1.5-7b_result.json```

Total: 4241, Correct: 2943, Accuracy: 69.39%, IMG-Accuracy: 67.87%

#!/bin/bash

python -m llava.eval.model_vqa_loader \
    --model-path liuhaotian/llava-v1.5-7b \
    --question-file ./playground/data/eval/MME/llava_mme.jsonl \
    --image-folder ./playground/data/eval/MME/MME_Benchmark_release_version \
    --answers-file ./playground/data/eval/MME/answers/llava-v1.5-7b.jsonl \
    --temperature 0 \
    --conv-mode vicuna_v1

cd ./playground/data/eval/MME

python convert_answer_to_mme.py --experiment llava-v1.5-7b

cd eval_tool

python calculation.py --results_dir answers/llava-v1.5-7b

=========== Perception ===========
total score: 1509.3788515406163 

         existence  score: 190.0
         count  score: 155.0
         position  score: 133.33333333333334
         color  score: 170.0
         posters  score: 147.61904761904762
         celebrity  score: 136.1764705882353
         scene  score: 158.0
         landmark  score: 162.25
         artwork  score: 119.5
         OCR  score: 137.5


=========== Cognition ===========
total score: 348.2142857142857 

         commonsense_reasoning  score: 110.71428571428571
         numerical_calculation  score: 70.0
         text_translation  score: 107.5
         code_reasoning  score: 60.0