# This script is for reproducing the result # GIVEN SFT models (4 epochs) # llama-3b-planner-bird-fft # llama-3b-validator-bird-fft # llama-3b-fixed-bird-fft mkdir data/llm_alignment/bird/ # ================================================================================================= # GEN ORPO data for planner # ================================================================================================= CUDA_VISIBLE_DEVICES=0 vllm serve llama-3b-bird-planner-fft/ --host 0.0.0.0 --port 8003 --dtype bfloat16 --max-model-len 4096 --disable-log-requests --served-model-name planner --gpu-memory-utilization 0.8 train_file=data/llm_alignment/llama-3b-planner-bird_train.jsonl rm $train_file PYTHONPATH=. python evaluate_end2end.py \ --input_file data/full_value_matching_schema_insight_bird_062024_with_evidence_train_text2sql.json \ --output_file $train_file \ --model-name llama --mode train \ --n_return 10 --temperature 1.0 --api_host http://localhost:8003 --only_planner DATA_DIR=data/llm_alignment/bird-p1-planner/ rm -r $DATA_DIR/ PYTHONPATH=. python llm_alignment/build_rl_data.py \ --input_file data/llm_alignment/orpo-llama-3b-planner-bird_train_beam_and_temperature.jsonl \ --gpt_planner_file data/multi-agents/planner/gpt-planner_combine_with_true_sql_bird_062024_with_evidence_train_combine.jsonl \ --output_planner_file $DATA_DIR/dpo-llama-3-end2end-bird_train_planner.jsonl \ --output_validator_select_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_select.jsonl \ --output_validator_condition_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_condition.jsonl \ --output_validator_join_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_join.jsonl \ --output_validator_order_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_order.jsonl \ --output_fixed_sql_file $DATA_DIR/dpo-llama-3-end2end-bird_train_fixed_sql.jsonl \ --n_select_chosens 2 # ================================================================================================= # GEN ORPO data for planner Iter 2 # ================================================================================================= train_file=data/llm_alignment/orpo-llama-3b-planner-bird_train_temperature.jsonl rm $train_file PYTHONPATH=. python evaluate_end2end.py \ --input_file data/full_value_matching_schema_insight_bird_062024_with_evidence_train_text2sql.json \ --output_file $train_file \ --model-name llama --mode train \ --n_return 10 --temperature 1.0 --api_host http://localhost:8003 --only_planner # PYTHONPATH=. python evaluate_end2end.py \ # --input_file data/full_value_matching_schema_insight_bird_062024_with_evidence_train_text2sql.json \ # --output_file $train_file \ # --model-name llama --mode train \ # --n_return 4 --use_beam_search --temperature 0.0 --api_host http://localhost:8003 --only_planner # DATA_DIR=data/llm_alignment/bird-p2-planner/ # rm -r $DATA_DIR/ # PYTHONPATH=. python llm_alignment/build_rl_data.py \ # --input_file data/llm_alignment/orpo-llama-3b-planner-bird_train_beam_and_temperature.jsonl \ # --gpt_planner_file data/multi-agents/planner/gpt-planner_combine_with_true_sql_bird_062024_with_evidence_train_combine.jsonl \ # --output_planner_file $DATA_DIR/dpo-llama-3-end2end-bird_train_planner.jsonl \ # --output_validator_select_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_select.jsonl \ # --output_validator_condition_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_condition.jsonl \ # --output_validator_join_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_join.jsonl \ # --output_validator_order_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_order.jsonl \ # --output_fixed_sql_file $DATA_DIR/dpo-llama-3-end2end-bird_train_fixed_sql.jsonl \ # --n_select_chosens 2 DATA_DIR=data/llm_alignment/bird-p2-planner/ rm -r $DATA_DIR/ PYTHONPATH=. python llm_alignment/build_rl_data.py \ --input_file $train_file \ --gpt_planner_file data/multi-agents/planner/gpt-planner_combine_with_true_sql_bird_062024_with_evidence_train_combine.jsonl \ --output_planner_file $DATA_DIR/dpo-llama-3-end2end-bird_train_planner.jsonl \ --output_validator_select_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_select.jsonl \ --output_validator_condition_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_condition.jsonl \ --output_validator_join_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_join.jsonl \ --output_validator_order_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_order.jsonl \ --output_fixed_sql_file $DATA_DIR/dpo-llama-3-end2end-bird_train_fixed_sql.jsonl \ --n_select_chosens 2 # ITER 3 vllm serve orpo-llama-3b-iter-2-bird-planner/ --host 0.0.0.0 --port 8103 --dtype auto --max-model-len 4096 --disable-log-requests --served-model-name planner --gpu-memory-utilization 0.9 train_file=data/llm_alignment/orpo-llama-3b-iter-2-planner-bird_train_temperature.jsonl rm $train_file PYTHONPATH=. python evaluate_end2end.py \ --input_file data/full_value_matching_schema_insight_bird_062024_with_evidence_train_text2sql.json \ --output_file $train_file \ --model-name llama --mode train \ --n_return 10 --temperature 1.0 --api_host http://localhost:8003 --only_planner DATA_DIR=data/llm_alignment/bird-p3-planner/ rm -r $DATA_DIR/ PYTHONPATH=. python llm_alignment/build_rl_data.py \ --input_file $train_file \ --gpt_planner_file data/multi-agents/planner/gpt-planner_combine_with_true_sql_bird_062024_with_evidence_train_combine.jsonl \ --output_planner_file $DATA_DIR/dpo-llama-3-end2end-bird_train_planner.jsonl \ --output_validator_select_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_select.jsonl \ --output_validator_condition_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_condition.jsonl \ --output_validator_join_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_join.jsonl \ --output_validator_order_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_order.jsonl \ --output_fixed_sql_file $DATA_DIR/dpo-llama-3-end2end-bird_train_fixed_sql.jsonl \ --n_select_chosens 2 scp -r $DATA_DIR grif:~/codes/data/llm_alignment/ # GEN ORPO FOR FIX # train_file=data/llm_alignment/orpo-llama-3b-iter-2-planner-bird_train_greedy.jsonl # PYTHONPATH=. python evaluate_end2end.py \ # --input_file data/full_value_matching_schema_insight_bird_062024_with_evidence_train_text2sql.json \ # --output_file $train_file \ # --model-name llama --mode train \ # --n_return 1 --temperature 0.0 --api_host http://192.168.1.108:8003 --only_planner python jsonl2json.py --jsonl-file data/llm_alignment/orpo-llama-3b-iter-2-planner-bird_train_greedy.jsonl PYTHONPATH=. python evaluate_end2end.py \ --input_file data/llm_alignment/orpo-llama-3b-iter-2-planner-bird_train_greedy.json \ --output_file data/llm_alignment/orpo-llama-3b-iter-2-validator-bird_train_greedy.jsonl\ --model-name llama --mode train \ --n_return 1 --temperature 0.0 --api_host http://localhost:8003 --skip_planner --skip_fix --skip_validator_order python jsonl2json.py --jsonl-file data/llm_alignment/orpo-llama-3b-iter-2-validator-bird_train_greedy.jsonl PYTHONPATH=. python evaluate_end2end.py \ --input_file data/llm_alignment/orpo-llama-3b-iter-2-validator-bird_train_greedy.json \ --output_file data/llm_alignment/orpo-llama-3b-iter-2-fix-bird_train_sampling.jsonl\ --model-name llama --mode train \ --n_return 10 --temperature 1.0 --api_host http://localhost:8003 --skip_planner --skip_validator DATA_DIR=data/llm_alignment/bird-p1-fix/ rm -r $DATA_DIR PYTHONPATH=. python llm_alignment/build_rl_data.py \ --input_file data/llm_alignment/orpo-llama-3b-iter-2-fix-bird_train_sampling.jsonl \ --output_planner_file $DATA_DIR/dpo-llama-3-end2end-bird_train_planner.jsonl \ --output_validator_select_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_select.jsonl \ --output_validator_condition_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_condition.jsonl \ --output_validator_join_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_join.jsonl \ --output_validator_order_file $DATA_DIR/dpo-llama-3-end2end-bird_train_validator_order.jsonl \ --output_fixed_sql_file $DATA_DIR/dpo-llama-3-end2end-bird_train_fixed_sql.jsonl # ================================================================================================= # GEN ORPO data for Validator # ================================================================================================= CUDA_VISIBLE_DEVICES=0 vllm serve llama-3b-bird-validator-fft/ --host 0.0.0.0 --port 8004 --dtype bfloat16 --max-model-len 4096 --disable-log-requests --served-model-name validator --gpu-memory-utilization 0.8 CUDA_VISIBLE_DEVICES=1 vllm serve llama-3b-bird-fixed-fft-follow-validation/ --host 0.0.0.0 --port 8005 --dtype bfloat16 --max-model-len 4096 --disable-log-requests --served-model-name fixed --gpu-memory-utilization 0.8 python jsonl2json.py --jsonl-file data/llm_alignment/bird/llama-3b-planner-bird_train.jsonl python jsonl2json.py --jsonl-file data/llm_alignment/bird/llama-3b-planner-bird_dev.jsonl rm data/llm_alignment/bird/llama-3b-validator-bird_train.jsonl PYTHONPATH=. python evaluate_end2end.py \ --input_file data/llm_alignment/bird/llama-3b-planner-bird_train.json \ --output_file data/llm_alignment/bird/llama-3b-validator-bird_train.jsonl \ --model-name llama --mode train \ --n_return 6 --temperature 0.7 --api_host http://localhost:8003 --skip_planner --skip_fix python jsonl2json.py --jsonl-file data/llm_alignment/bird/llama-3b-validator-bird_train.jsonl rm data/llm_alignment/bird/llama-3b-fix-greedy-bird_train.jsonl PYTHONPATH=. python evaluate_end2end.py \ --input_file data/llm_alignment/bird/llama-3b-validator-bird_train.json \ --output_file data/llm_alignment/bird/llama-3b-fix-greedy-bird_train.jsonl \ --model-name llama --mode train \ --n_return 1 --temperature 0.0 --api_host http://localhost:8003 --skip_planner --skip_validator rm -r data/llm_alignment/bird/bird-p1-validator/ PYTHONPATH=. python llm_alignment/build_rl_data.py \ --input_file data/llm_alignment/bird/llama-3b-fix-greedy-bird_train.jsonl \ --output_planner_file data/llm_alignment/bird/bird-p1-validator/dpo-llama-3-end2end-bird_train_planner.jsonl \ --output_validator_select_file data/llm_alignment/bird/bird-p1-validator/dpo-llama-3-end2end-bird_train_validator_select.jsonl \ --output_validator_condition_file data/llm_alignment/bird/bird-p1-validator/dpo-llama-3-end2end-bird_train_validator_condition.jsonl \ --output_validator_join_file data/llm_alignment/bird/bird-p1-validator/dpo-llama-3-end2end-bird_train_validator_join.jsonl \ --output_validator_order_file data/llm_alignment/bird/bird-p1-validator/dpo-llama-3-end2end-bird_train_validator_order.jsonl \ --output_fixed_sql_file data/llm_alignment/bird/bird-p1-validator/dpo-llama-3-end2end-bird_train_fixed_sql.jsonl # Use Feedback Editor Agent rm processed_results.json python modify_feedbacks.py --pred_file data/llm_alignment/bird/llama-3b-fix-greedy-bird_train.jsonl rm -r data/llm_alignment/bird/bird-p1-validator-modify PYTHONPATH=. python llm_alignment/build_rl_data.py \ --input_file data/llm_alignment/bird/llama-3b-fix-greedy-bird_train_progress.jsonl \ --output_planner_file data/llm_alignment/bird/bird-p1-validator-modify/dpo-llama-3-end2end-bird_train_planner.jsonl \ --output_validator_select_file data/llm_alignment/bird/bird-p1-validator-modify/dpo-llama-3-end2end-bird_train_validator_select.jsonl \ --output_validator_condition_file data/llm_alignment/bird/bird-p1-validator-modify/dpo-llama-3-end2end-bird_train_validator_condition.jsonl \ --output_validator_join_file data/llm_alignment/bird/bird-p1-validator-modify/dpo-llama-3-end2end-bird_train_validator_join.jsonl \ --output_validator_order_file data/llm_alignment/bird/bird-p1-validator-modify/dpo-llama-3-end2end-bird_train_validator_order.jsonl \ --output_fixed_sql_file data/llm_alignment/bird/bird-p1-validator-modify/dpo-llama-3-end2end-bird_train_fixed_sql.jsonl # ================================================================================================= # GEN ORPO data for FIX # ================================================================================================= CUDA_VISIBLE_DEVICES=1 vllm serve llama-3b-bird-fixed-fft-follow-validation/ --host 0.0.0.0 --port 8005 --dtype bfloat16 --max-model-len 4096 --disable-log-requests --served-model-name fixed --gpu-memory-utilization 0.8 rm data/llm_alignment/bird/llama-3b-validator-greedy-bird_train.jsonl PYTHONPATH=. python evaluate_end2end.py \ --input_file data/llm_alignment/bird/llama-3b-planner-bird_train.json \ --output_file data/llm_alignment/bird/llama-3b-validator-greedy-bird_train.jsonl \ --model-name llama --mode train \ --n_return 1 --temperature 0.0 --api_host http://localhost:8003 --skip_planner --skip_fix python jsonl2json.py --jsonl-file data/llm_alignment/bird/llama-3b-validator-greedy-bird_train.jsonl rm data/llm_alignment/bird/llama-3b-fix-sampling-bird_train.jsonl PYTHONPATH=. python evaluate_end2end.py \ --input_file data/llm_alignment/bird/llama-3b-validator-greedy-bird_train.json \ --output_file data/llm_alignment/bird/llama-3b-fix-sampling-bird_train.jsonl \ --model-name llama --mode train \ --n_return 10 --temperature 0.7 --api_host http://localhost:8003 --skip_planner --skip_validator rm -r data/llm_alignment/bird/bird-p1-fix/ PYTHONPATH=. python llm_alignment/build_rl_data.py \ --input_file data/llm_alignment/bird/llama-3b-fix-sampling-bird_train.jsonl \ --output_planner_file data/llm_alignment/bird/bird-p1-fix/dpo-llama-3-end2end-bird_train_planner.jsonl \ --output_validator_select_file data/llm_alignment/bird/bird-p1-fix/dpo-llama-3-end2end-bird_train_validator_select.jsonl \ --output_validator_condition_file data/llm_alignment/bird/bird-p1-fix/dpo-llama-3-end2end-bird_train_validator_condition.jsonl \ --output_validator_join_file data/llm_alignment/bird/bird-p1-fix/dpo-llama-3-end2end-bird_train_validator_join.jsonl \ --output_validator_order_file data/llm_alignment/bird/bird-p1-fix/dpo-llama-3-end2end-bird_train_validator_order.jsonl \ --output_fixed_sql_file data/llm_alignment/bird/bird-p1-fix/dpo-llama-3-end2end-bird_train_fixed_sql.jsonl \ --enable_advanced_fix_agent