| # Stage 2b: de-bias stage 2 with short-answer VQA data (continues from stage-2 LoRA + projector), then evaluate. | |
| set -uo pipefail | |
| cd "$HOME/atlas" | |
| export PYTHONUNBUFFERED=1 HF_HOME="$HOME/.cache/huggingface" PATH="$HOME/.local/bin:$PATH" | |
| echo "=== [1/4] build de-bias mix ===" | |
| python3 prepare_mix.py || exit 11 | |
| export HF_HUB_OFFLINE=1 | |
| TRAIN_ENV="INSTRUCT_JSON=$HOME/data/llava_instruct/mix_debias.json INIT_FROM=$HOME/checkpoints/stage2/latest.pt LORA_LR=${LORA_LR:-1e-4} PROJ_LR=${PROJ_LR:-1e-5} HELDOUT=200" | |
| echo "=== [2/4] smoke test (10 steps) ===" | |
| rm -rf "$HOME/checkpoints/stage2b_smoke" | |
| env $TRAIN_ENV MAX_STEPS=10 LOG_EVERY=2 SAVE_EVERY=1000000 CKPT_DIR="$HOME/checkpoints/stage2b_smoke" python3 train_stage2.py || { echo "SMOKE TEST FAILED"; exit 12; } | |
| echo "=== [3/4] full stage-2b run ===" | |
| env $TRAIN_ENV CKPT_DIR="$HOME/checkpoints/stage2b" python3 train_stage2.py || { echo "TRAINING FAILED"; exit 13; } | |
| echo "=== [4/4] evaluation (same held-out set and POPE as stage 2) ===" | |
| env -u INSTRUCT_JSON -u INIT_FROM -u HELDOUT CKPT_DIR="$HOME/checkpoints/stage2b" EVAL_DIR="$HOME/eval/stage2b" python3 eval_stage2.py || { echo "EVAL FAILED"; exit 14; } | |
| cat "$HOME/eval/stage2b/report.md" | |
| echo "=== run_stage2b.sh finished ===" | |