theia / code /COMMANDS_QUICKSTART.txt
overmach's picture
Camera-ready release
148072e verified
Raw History Blame Contribute Delete
5.21 kB
COMMANDS QUICKSTART (repo/)
This file provides copy/paste command combinations for the main entrypoints:
- VLM_finetune/run_ablation_sequential_ft.py
- VLM_inference/run_ft_eval_ablation.py
- VLM_inference/test_base_models/run_ft_eval_update.py
Notes
- Replace <USER> in paths with your username.
- These scripts assume your datasets/adapters follow the folder conventions described in their headers.
- All commands are meant to be run from /users/<USER>/repo unless you prefer absolute paths.
1) Sequential fine-tuning launcher
Script: VLM_finetune/run_ablation_sequential_ft.py
A. Dry-run: list all planned runs (no training launched)
python VLM_finetune/run_ablation_sequential_ft.py --dry_run
B. Train a single task at a single percentage
python VLM_finetune/run_ablation_sequential_ft.py --task a1 --perc 100 --
C. Train multiple tasks and multiple dataset percentages
python VLM_finetune/run_ablation_sequential_ft.py --tasks a1,b1,c1 --percs 25,50,75,100 --
D. Overwrite existing outputs (dangerous: deletes output folders)
python VLM_finetune/run_ablation_sequential_ft.py --task a1 --perc 25 --overwrite --
E. Change splits root / output root / HF cache location
python VLM_finetune/run_ablation_sequential_ft.py \
--splits_root /users/<USER>/ablation_study_dataset/ablation_splits \
--output_root /mydata/<USER>/ablation_study_outputs/finetune_lora \
--hf_home /mydata/<USER>/VLM_cache \
--tasks a1,b1 --percs 50,100 \
--
F. Forward extra FT.py arguments (everything after "--" is passed through)
python VLM_finetune/run_ablation_sequential_ft.py --task a1 --perc 25 -- \
--num_train_epochs 1 \
--learning_rate 2e-5 \
--per_device_train_batch_size 1 \
--gradient_accumulation_steps 8 \
2) Ablation evaluation runner (evaluate many adapters)
Script: VLM_inference/run_ft_eval_ablation.py
A. Evaluate ALL discovered (task, perc) splits under splits-root
python VLM_inference/run_ft_eval_ablation.py \
--splits-root /users/<USER>/ablation_study_dataset_c1/ablation_splits \
--adapter-root /mydata/<USER>/ablation_study_outputs/finetune_lora \
--cache-dir /mydata/<USER>/VLM_cache
B. Evaluate a subset of tasks and percs
python VLM_inference/run_ft_eval_ablation.py --tasks a1,b1 --percs 25,50,100
C. Evaluate only a few samples per split (fast smoke test)
python VLM_inference/run_ft_eval_ablation.py --tasks a1 --percs 100 --num-samples 20
D. Skip missing adapters (useful while training is still running)
python VLM_inference/run_ft_eval_ablation.py --skip-missing-adapter
E. Evaluate BASE model only (ignore adapters)
python VLM_inference/run_ft_eval_ablation.py --tasks a1 --percs 100 --no-adapter
F. Evaluate exactly one adapter path (ignore task/perc discovery)
python VLM_inference/run_ft_eval_ablation.py \
--adapter /mydata/<USER>/ablation_study_outputs/finetune_lora/task_a1/perc_100
G. Choice-task robustness options (avoid whitespace-only outputs)
python VLM_inference/run_ft_eval_ablation.py --tasks a1 --percs 100 \
--retry-empty-choice 2 \
--force-choice-letter-fallback
H. Box-task generation settings (sampling / beams)
python VLM_inference/run_ft_eval_ablation.py --tasks b2 --percs 100 \
--max-new-tokens-boxes 4096 \
--boxes-num-beams 3
I. Dry-run (show planned runs but do not execute inference)
python VLM_inference/run_ft_eval_ablation.py --dry-run
3) Single-run evaluator (one task, optional circuit, optional adapter)
Script: VLM_inference/test_base_models/run_ft_eval_update.py
A. Evaluate one task across ALL circuits (default sampling logic)
python VLM_inference/test_base_models/run_ft_eval_update.py --task a1 --num-samples 200
B. Evaluate one task for one circuit
python VLM_inference/test_base_models/run_ft_eval_update.py --task b2 --circuit a --num-samples 200
C. Evaluate an ablation task (a1/b1/c1) using its ablation split JSON
python VLM_inference/test_base_models/run_ft_eval_update.py --task a1 --num-samples -1
D. Evaluate with a LoRA adapter
python VLM_inference/test_base_models/run_ft_eval_update.py \
--task a1 --num-samples 200 \
--adapter /mydata/<USER>/ablation_study_outputs/finetune_lora/task_a1/perc_100
E. Evaluate a different base model (no adapter)
python VLM_inference/test_base_models/run_ft_eval_update.py \
--task a1 --num-samples 50 \
--model meta-llama/Llama-3.2-11B-Vision
F. Set cache dir for model/tokenizer downloads
python VLM_inference/test_base_models/run_ft_eval_update.py --task a1 --cache-dir /mydata/<USER>/VLM_cache
G. Debug Hugging Face auth without running full inference
python VLM_inference/test_base_models/run_ft_eval_update.py \
--task a1 --num-samples 5 --dry-run-hf \
--model meta-llama/Llama-3.2-11B-Vision \
--hf-token-file /users/<USER>/.secrets/hf_token.env
H. Force single GPU (avoid some device_map='auto' issues)
python VLM_inference/test_base_models/run_ft_eval_update.py --task a1 --force-single-gpu
I. Change output directory and run tag
python VLM_inference/test_base_models/run_ft_eval_update.py \
--task a1 --num-samples 200 \
--out-dir /users/<USER>/VLM/inference_eval/outputs \
--run-tag my_experiment_001