vital2TS — VITAL student / teacher / data / CuPy bundle

Artifacts for reproducing VITAL (vis-sigmoid OPD + ROUGE-L gap k3) training with Qwen3-VL.

Layout

Path	Contents	Source
student_2b_sft/	Qwen3-VL-2B SFT merged HF	MVP/checkpoints/.../global_step_250_merged_hf (~4.6G)
teacher_8b_gspo/	Qwen3-VL-8B GSPO merged HF	.../global_step_1500/actor_merged_hf (~17G)
data/mimiccxr_verl_4k_imgcap1500/train_grpo_teacher_rougel.parquet	Train parquet with teacher ROUGE-L cache	same-named local file
data/mimiccxr_verl_4k_imgcap1500/val_grpo.parquet	Val parquet	same-named local file
deps/cupy-cuda12x-13.6.0/	Bundled CuPy 13.6.0 (cuda12x) for scripts that set CUPY_PREFIX / PYTHONPATH	Qwen3-VL-8B-Instruct/.deps/cupy-cuda12x-13.6.0 (~225M)

Env hooks (match training scripts)

export STUDENT_MODEL=$PWD/student_2b_sft
export TEACHER_MODEL=$PWD/teacher_8b_gspo
export TRAIN_FILE=$PWD/data/mimiccxr_verl_4k_imgcap1500/train_grpo_teacher_rougel.parquet
export TEST_FILE=$PWD/data/mimiccxr_verl_4k_imgcap1500/val_grpo.parquet
export CUPY_PREFIX=$PWD/deps/cupy-cuda12x-13.6.0
export PYTHONPATH="${CUPY_PREFIX}${PYTHONPATH:+:$PYTHONPATH}"

Code / split lists: see companion repo VITAL-for-Medical-Diagnosis (does not ship these weights).

Note on CuPy

Training sbatch defaults expect .deps/cupy-cuda12x-13.6.0. A newer CuPy 14 in conda does not satisfy that path check; use this bundled tree or point CUPY_PREFIX at it.