thanhdath/mats-sql-bundle

MATS-SQL Bundle — Paper-Format Pipeline (iter1)

Full MATS multi-agent Text2SQL pipeline reproduced on BIRD-bench. Validators use the paper's Feedback: + Conclude: correct/incorrect. format (not wrapper-tag).

Final BIRD-dev numbers (1524/1534 questions with usable db_path)

Config	planner@1 (T=0)	pipeline@1 (T=0)	oracle pass@8	trained selector EX
PLANNER-only	51.54%	51.54%	70.80%	—
SFT-VF (paper validators)	51.48%	52.20%	71.65%	59.91%
ORPO iter1 COLLAB	51.08%	51.81%	71.19%	59.97%
ORPO iter1 INDEP	51.74%	52.59%	71.95%	60.31%

All three pipelines converge to ~60% selector EX. Validator-internal ORPO gains (COLLAB reward-acc +10/+17pp over INDEP) do not translate to pipeline EX.

Layout

models/
  planner-3B-sft/                          # Qwen2.5-Coder-3B, planner SFT
  planner-3B-orpo-iter1/                   # planner ORPO iter1 (Alg.1)
  validator-sel-1B-sft-paper/              # Llama-3.2-1B, val-sel SFT (paper format)
  validator-sel-1B-orpo-iter1-collab-paper/
  validator-sel-1B-orpo-iter1-indep-paper/
  validator-cond-1B-sft-paper/             # Llama-3.2-1B, val-cond SFT (paper format)
  validator-cond-1B-orpo-iter1-collab-paper/
  validator-cond-1B-orpo-iter1-indep-paper/
  fixer-1B-sft/                            # Llama-3.2-1B, fixer SFT
  fixer-1B-orpo-iter1/                     # fixer ORPO iter1
  selector-3B-sft/                         # Qwen2.5-Coder-3B, selector SFT

data/
  hf_val_sel_paper_v1/                     # paper-format val-sel SFT data (8890 train + 468 test)
  hf_val_cond_paper_v1/                    # paper-format val-cond SFT data
  hf_orpo_val_{sel,cond}_paper_iter1_{collab,indep}/   # ORPO iter1 pairs
  planner_3B_greedy_bird_train.jsonl       # planner-3B greedy preds on BIRD-train (9360 q, 57.6% correct)

eval_results/
  paper_{SFT_VF,COLLAB_par,INDEP_par}_passAt8_bird_dev.jsonl  # K=8 T=1.0 full rollouts
  paper_greedy_{PLANNER_ONLY,SFT_VF,COLLAB,INDEP}_passAt1_bird_dev.jsonl  # T=0 K=1 rollouts

scripts/      # pipeline + builders + trainer + selector EX
recipes/iter1-paper/  # ORPO recipes for the 4 validator variants

HANDOFF_SELECTOR_TASK.md  # next-step doc: lift selector EX from 60% to 67%

Validator training reward accuracies (eval_dpo on test split)

Model	eval_loss	eval_rewards/accuracies
validator-sel-1B-orpo-iter1-collab-paper	0.174	69.7%
validator-sel-1B-orpo-iter1-indep-paper	0.210	59.7%
validator-cond-1B-orpo-iter1-collab-paper	0.148	89.7%
validator-cond-1B-orpo-iter1-indep-paper	0.163	72.0%

COLLAB > INDEP by +10pp (sel) and +17.7pp (cond) at the chosen-vs-rejected level, but the gain doesn't propagate to end-to-end EX. See HANDOFF_SELECTOR_TASK.md.

Hyperparameters

• SFT (validator/fixer/selector): lr=2e-5 cosine, 2 epochs, bf16, completion-only loss.
• ORPO (paper §4.3): β=0.1, lr=8e-6 (1e-6 for val-sel-indep due to NaN), max_steps=400, max_grad_norm=0.5.
• Rollout (run_pipeline_rollouts.py): K=8 candidates per question, T=1.0, top_p=0.9.
• Fixer gated by planner_exec_ok=False.

Reproduce

# 1. Generate paper-format validator SFT data (uses Qwen-2.5-72B as teacher)
python scripts/gen_planner_preds_for_validator.py --out data/planner_3B_greedy_bird_train.jsonl  # ~15min
python scripts/gen_validator_sft_qwen72b.py --input data/planner_3B_greedy_bird_train.jsonl     # ~7h

# 2. Train paper-format validators
python scripts/train_sft_completion_only.py --base meta-llama/Llama-3.2-1B-Instruct \
    --data data/hf_val_sel_paper_v1 --chat_format llama3 --epochs 2 --lr 2e-5

# 3. Generate ORPO iter1 data (collab + indep)
python scripts/build_orpo_data.py --agent validator_sel --mode collab --K 4 --temperature 1.0 --max_questions 2000

# 4. ORPO train (per recipe in recipes/iter1-paper/)
accelerate launch alignment-handbook/scripts/run_orpo.py recipes/iter1-paper/orpo-val-sel-collab-paper.yaml

# 5. K=8 rollout eval on BIRD-dev
python scripts/run_pipeline_rollouts.py --input data/sft_bird_with_evidence_dev_text2sql.json \
    --K 8 --temperature 1.0 ...

# 6. Selector EX
python scripts/compute_bestofn_with_selector.py eval_results/paper_*_passAt8_bird_dev.jsonl

Open problem — see HANDOFF_SELECTOR_TASK.md

Selector EX is saturated at ~60% across all three validator regimes. Target is ≥67%. Oracle pass@8 ≈ 72% gives 7pp headroom; selector picks 60/72 = 83% of available correct. Next agent should focus on lifting selector pick-rate (or oracle@8 ceiling), not validators.