QWENVL-PRIVATE/make_stage3_samples.sh

# Common paths
DATA_ROOT="/root/epfs/sproject/datasets"
STAGE3_RESULTS="${DATA_ROOT}/stage3results"

# Ensure the output directory exists
mkdir -p "${STAGE3_RESULTS}"

# 1. ChartQA 28k
python analysis/making_gated_training_samples.py \
    --model-path /root/public/Qwen/Qwen2.5-VL-3B-Instruct \
    --question-file "${DATA_ROOT}/chartqa_28k_qa.jsonl" \
    --image-folder "${DATA_ROOT}" \
    --dst-save-path "${STAGE3_RESULTS}/gated_chartqa28k_qwen2.5vl_3b.pkl" \
    --gpu-ids 0,1,2,3

# 2. Visual CoT LLaVA 68k
python analysis/making_gated_training_samples.py \
    --model-path /root/public/Qwen/Qwen2.5-VL-3B-Instruct \
    --question-file "${DATA_ROOT}/visual_cot_llava_subset68k.jsonl" \
    --image-folder "${DATA_ROOT}" \
    --dst-save-path "${STAGE3_RESULTS}/gated_vcot_llava68k_qwen2.5vl_3b.pkl" \
    --gpu-ids 0,1,2,3

# 3. Visual CoT DocVQA 33k
python analysis/making_gated_training_samples.py \
    --model-path /root/public/Qwen/Qwen2.5-VL-3B-Instruct \
    --question-file "${DATA_ROOT}/visual_cot_docvqa_subset33k.jsonl" \
    --image-folder "${DATA_ROOT}" \
    --dst-save-path "${STAGE3_RESULTS}/gated_vcot_docvqa33k_qwen2.5vl_3b.pkl" \
    --gpu-ids 0,1,2,3