mats-sql-bundle / code /llm_alignment /build_rl_data_collaborative.py
thanhdath's picture
Push code: scripts, slurm sbatch, recipes, utils (v3 + selector series)
778d47d verified
Raw
History Blame Contribute Delete
9 kB
"""
Build preference-pair data from 3-stage pipeline rollouts (output of run_pipeline_rollouts.py).
For three trainable agents (planner / validator / fixer), emits preference-pair files for
the INDEPENDENT and COLLABORATIVE training schemes:
PLANNER:
- independent: chosen/rejected by is_correct(planner_sql) (LOCAL label)
- collaborative: chosen/rejected by is_correct(fixed_sql) (TRAJECTORY label)
VALIDATOR (free-text critique — NO natural local label):
- independent: SKIPPED (cannot be constructed without an external teacher;
the methodology section explains this is the structural
reason MATS originally depended on GPT-4o-mini.)
- collaborative: chosen/rejected by is_correct(fixed_sql) (TRAJECTORY label)
FIXER (terminal stage):
- independent: chosen/rejected by is_correct(fixed_sql)
- collaborative: chosen/rejected by is_correct(fixed_sql) (same — terminal)
The (e) vs (b)/(d) ablation: the methodological gap is the validator-collab line.
Without collaborative training, the validator pair set is empty.
Usage:
python llm_alignment/build_rl_data_collaborative.py \\
--rollouts data/rollouts/bird_train_3stage_K4.jsonl \\
--output_dir data/llm_alignment/collab/
"""
import argparse
import json
import os
import random
import sys
def build_pairs(samples, completion_field, label_field, prompt_field, share_prompt=False):
"""
For each question, pair winners vs losers.
When `share_prompt=True` (planner case): chosen and rejected must come from trajectories
sharing the same prompt (standard ORPO interface).
When `share_prompt=False` (validator/fixer case): pairs are formed across the question;
the prompt of the *winning* trajectory is used for both chosen and rejected. This is the
methodologically simplest tractable formulation when intermediate-agent outputs are
near-identical within a fixed upstream context (templated SFT data + small T=0.7 effect).
"""
pairs = []
for s in samples:
if share_prompt:
prompt_to_traj = {}
for t in s.get("trajectories", []):
p = t.get(prompt_field)
if p is None:
continue
prompt_to_traj.setdefault(p, []).append(t)
buckets = list(prompt_to_traj.items())
else:
# One bucket per question; emit pairs across all trajectories.
ts_all = [t for t in s.get("trajectories", []) if t.get(prompt_field) is not None]
buckets = [(None, ts_all)] if ts_all else []
for _prompt_key, ts in buckets:
wins = [t for t in ts if label_field(t)]
losses = [t for t in ts if not label_field(t)]
if not wins or not losses:
continue
for w in wins[:2]:
for l in losses[:2]:
cw = completion_field(w)
cl = completion_field(l)
if not cw or not cl:
continue
if cw.strip() == cl.strip():
continue
# Use winning trajectory's prompt for both chosen and rejected
# (when share_prompt=False); this is the standard ORPO interface
# adaptation and is documented in the methodology section.
use_prompt = w.get(prompt_field) if not share_prompt else _prompt_key
if use_prompt is None:
continue
pairs.append({
"prompt": use_prompt,
"chosen": cw,
"rejected": cl,
"db_path": s.get("db_path"),
"question": s.get("question"),
"db_id": s.get("db_id"),
})
return pairs
def write_jsonl(path, rows):
os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
with open(path, "w") as f:
for r in rows:
f.write(json.dumps(r) + "\n")
print(f" wrote {len(rows):>7d} pairs → {path}")
def write_hf_dataset(out_dir, rows, train_frac=0.95):
from datasets import Dataset, DatasetDict
if not rows:
print(f" SKIP {out_dir} — no rows")
return
random.seed(42)
idxs = list(range(len(rows)))
random.shuffle(idxs)
n_train = max(1, int(len(rows) * train_frac))
train_rows = [rows[i] for i in idxs[:n_train]]
test_rows = [rows[i] for i in idxs[n_train:]] or [rows[-1]]
ds = DatasetDict({
"train_dpo": Dataset.from_list(train_rows),
"test_dpo": Dataset.from_list(test_rows),
})
if os.path.exists(out_dir):
import shutil
shutil.rmtree(out_dir)
os.makedirs(out_dir, exist_ok=True)
ds.save_to_disk(out_dir)
print(f" wrote HF DatasetDict (train={len(train_rows)}, test={len(test_rows)}) → {out_dir}")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--rollouts", required=True)
parser.add_argument("--output_dir", default="data/llm_alignment/collab/")
parser.add_argument("--no_hf", action="store_true")
args = parser.parse_args()
print(f"Loading {args.rollouts}...")
samples = []
with open(args.rollouts) as f:
for line in f:
line = line.strip()
if not line:
continue
samples.append(json.loads(line))
print(f" {len(samples)} samples")
# Stats
n_with_winloss = 0
n_traj = 0
for s in samples:
traj = s.get("trajectories", [])
n_traj += len(traj)
wins = sum(1 for t in traj if t.get("is_fixed_correct"))
losses = sum(1 for t in traj if not t.get("is_fixed_correct"))
if wins > 0 and losses > 0:
n_with_winloss += 1
print(f" total trajectories: {n_traj}")
print(f" questions with both win+loss: {n_with_winloss} ({100*n_with_winloss/max(len(samples),1):.1f}%)")
# Planner — 2 variants (indep, collab); shared-prompt within trajectory group
print("\n[planner] building pairs (share_prompt=True — planner_prompt is identical across rollouts of same question)...")
indep_planner = build_pairs(
samples,
completion_field=lambda t: t.get("planner_output"),
label_field=lambda t: t.get("is_planner_correct", False),
prompt_field="planner_prompt",
share_prompt=True,
)
collab_planner = build_pairs(
samples,
completion_field=lambda t: t.get("planner_output"),
label_field=lambda t: t.get("is_fixed_correct", False),
prompt_field="planner_prompt",
share_prompt=True,
)
# Validator — collab only; cross-trajectory pairing
# (validator_prompt depends on planner_sql which differs across rollouts)
print("\n[validator] building COLLABORATIVE pairs (cross-trajectory; uses winning-traj prompt)...")
collab_validator = build_pairs(
samples,
completion_field=lambda t: t.get("validator_output"),
label_field=lambda t: t.get("is_fixed_correct", False),
prompt_field="validator_prompt",
share_prompt=False,
)
# Fixer — terminal; cross-trajectory pairing as well
print("\n[fixer] building pairs (cross-trajectory; uses winning-traj prompt)...")
fixer_pairs = build_pairs(
samples,
completion_field=lambda t: t.get("fixer_output"),
label_field=lambda t: t.get("is_fixed_correct", False),
prompt_field="fixer_prompt",
share_prompt=False,
)
out = args.output_dir
# JSONL outputs
write_jsonl(os.path.join(out, "planner_pairs_independent.jsonl"), indep_planner)
write_jsonl(os.path.join(out, "planner_pairs_collaborative.jsonl"), collab_planner)
write_jsonl(os.path.join(out, "validator_pairs_collaborative.jsonl"), collab_validator)
write_jsonl(os.path.join(out, "fixer_pairs_shared.jsonl"), fixer_pairs)
# HF DatasetDict outputs
if not args.no_hf:
write_hf_dataset(os.path.join(out, "hf_planner_independent"), indep_planner)
write_hf_dataset(os.path.join(out, "hf_planner_collaborative"), collab_planner)
write_hf_dataset(os.path.join(out, "hf_validator_collaborative"), collab_validator)
write_hf_dataset(os.path.join(out, "hf_fixer_shared"), fixer_pairs)
# Summary
print("\n=== Summary ===")
print(f" Planner pairs — indep: {len(indep_planner):>5d} | collab: {len(collab_planner):>5d}")
print(f" Validator pairs — indep: skipped (needs GPT) | collab: {len(collab_validator):>5d}")
print(f" Fixer pairs — shared: {len(fixer_pairs):>5d}")
print()
print(" Validator-collab is the methodologically novel pair set: it is GPT-free")
print(" AND the only pair set the validator can be aligned on without an external teacher.")
if __name__ == "__main__":
main()