File size: 8,457 Bytes
778d47d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 | """
Validator v4 ORPO data builder.
SFT trains on single completions. ORPO adds a contrastive signal:
- wrong SQL: chosen = "INCORRECT: [critique]", rejected = "None"
→ model learns: "don't stay silent on wrong SQL"
- correct SQL: chosen = "None", rejected = "INCORRECT: [critique]"
→ model learns: "don't falsely flag correct SQL"
Each example becomes ONE ORPO pair (prompt, chosen, rejected).
One dataset handles both sel (SELECT critique) and cond (CONDITION critique)
by creating two rows per trajectory — one per validator role.
Output: data/hf_validator_v4_orpo/{train_dpo, test_dpo}
columns: prompt, chosen, rejected, question, db_id
"""
import json, os, re, random, sqlite3
from datasets import Dataset, DatasetDict
ROOT = "/weka/s225250685/mats-tist"
os.chdir(ROOT)
SRC_PATHS = [
"data/rollouts/scaleup_bird_train_2stage_K4.jsonl",
"data/rollouts/bird_train_3stage_K4.jsonl",
"data/rollouts/iter2_bird_train_3stage_K8.jsonl",
]
OUT_DIR = "data/hf_validator_v4_orpo"
SEL_INSTR = ("You are a SQL SELECT-clause critique agent. Output ONE critique section "
"<select>...</select> analysing the SELECT clause of the SQL query below; "
"do NOT output any SQL. Use 'None' if the SELECT clause looks correct.")
COND_INSTR = ("You are a SQL CONDITION critique agent. Output ONE critique section "
"<condition>...</condition> analysing the WHERE/HAVING/CASE-WHEN conditions "
"of the SQL query below; do NOT output any SQL. Use 'None' if the conditions look correct.")
NONE_SEL = "<select>\nSELECT.\nNone\n</select>"
NONE_COND = "<condition>\nCONDITION.\nNone\n</condition>"
def resolve_db(d):
p = d.get("db_path", "")
if p and os.path.exists(p): return p
db_id = d.get("db_id", "")
for tmpl in [f"data/train_databases/{db_id}/{db_id}.sqlite",
f"data/dev_databases/{db_id}/{db_id}.sqlite"]:
if os.path.exists(tmpl): return tmpl
return None
def exec_str(db_path, sql, n=5):
try:
conn = sqlite3.connect(db_path)
conn.text_factory = lambda b: b.decode(errors="ignore")
rows = conn.execute(sql).fetchmany(n)
conn.close()
return str(rows)[:300]
except Exception as e:
return f"Error: {str(e)[:150]}"
def safe_trunc(s, n=2800):
s = str(s or ""); return s if len(s) <= n else s[:n] + "..."
def gen_sel_critique(wrong, gold):
wl, gl = wrong.lower(), gold.lower()
issues = []
for agg in ["count(", "sum(", "avg(", "max(", "min("]:
if agg in gl and agg not in wl: issues.append(f"Missing {agg[:-1].upper()}")
elif agg in wl and agg not in gl: issues.append(f"Unexpected {agg[:-1].upper()}")
if "distinct" in gl and "distinct" not in wl: issues.append("Missing DISTINCT")
elif "distinct" in wl and "distinct" not in gl: issues.append("Unexpected DISTINCT")
gs, ws = gl.count("select")-1, wl.count("select")-1
if gs > ws: issues.append("Missing subquery")
d = ("INCORRECT: " + "; ".join(issues) + ".") if issues else \
"INCORRECT: SELECT clause returns wrong results."
return f"<select>\nSELECT.\n{d}\n</select>"
def gen_cond_critique(wrong, gold):
wl, gl = wrong.lower(), gold.lower()
issues = []
gj, wj = gl.count("join"), wl.count("join")
if gj > wj: issues.append(f"Missing JOIN")
elif wj > gj: issues.append(f"Extra JOIN")
if ("group by" in gl) != ("group by" in wl): issues.append("GROUP BY mismatch")
if "having" in gl and "having" not in wl: issues.append("Missing HAVING")
if ("limit" in gl) != ("limit" in wl): issues.append("LIMIT mismatch")
d = ("INCORRECT: " + "; ".join(issues) + ".") if issues else \
"INCORRECT: WHERE/HAVING conditions return wrong results."
return f"<condition>\nCONDITION.\n{d}\n</condition>"
def build_prompt(instr, schema, question, evidence, sql, exec_result):
# Field labels must match VALIDATOR_SEL_HEADER/COND_HEADER + VALIDATOR_PROMPT_BODY
# in run_pipeline_rollouts.py exactly, so the trained model generalises to inference.
return (instr + "\n\ndatabase schema:\n" + schema +
"\n\nQuestion: " + question +
"\nExternal knowledge: " + (evidence or "None") +
"\n\nGenerated SQL query: " + sql +
"\n\nExecution response:\n" + exec_result + "\n\n")
def main():
rng = random.Random(42)
pairs = [] # each: {prompt, chosen, rejected, question, db_id}
seen = set()
for src in SRC_PATHS:
if not os.path.exists(src):
print(f"skip {src}"); continue
n_wrong = n_correct = 0
with open(src) as f:
for line in f:
line = line.strip()
if not line: continue
d = json.loads(line)
db_path = resolve_db(d)
if not db_path: continue
schema = safe_trunc(str(d.get("schema", "")), 2500)
question = d.get("question", "")
evidence = d.get("evidence", "") or "None"
gold_sql = (d.get("sql") or "").strip()
for t in d.get("trajectories", []):
sql = (t.get("planner_sql") or "").strip()
if not sql: continue
exec_ok = bool(t.get("planner_exec_ok", True))
if not exec_ok: continue # only exec_ok=True cases
correct = bool(t.get("is_planner_correct") or t.get("is_fixed_correct"))
key = (hash(question), sql[:60])
if key in seen: continue
seen.add(key)
es = exec_str(db_path, sql)
if not correct and gold_sql:
# WRONG SQL — teach validator to flag it
sel_crit = gen_sel_critique(sql, gold_sql)
cond_crit = gen_cond_critique(sql, gold_sql)
for instr, tag, chosen_crit, none_crit in [
(SEL_INSTR, "select", sel_crit, NONE_SEL),
(COND_INSTR, "condition", cond_crit, NONE_COND),
]:
prompt = build_prompt(instr, schema, question, evidence, sql, es)
pairs.append({"prompt": prompt, "chosen": chosen_crit,
"rejected": none_crit, # rejected = staying silent
"question": question, "db_id": d.get("db_id", ""),
"role": tag, "label": "wrong"})
n_wrong += 1
elif correct:
# CORRECT SQL — teach validator NOT to flag it
# Rejected = a plausible-looking but wrong critique
for instr, tag, none_crit, fake_critique in [
(SEL_INSTR, "select", NONE_SEL,
"<select>\nSELECT.\nINCORRECT: SELECT clause returns wrong results.\n</select>"),
(COND_INSTR, "condition", NONE_COND,
"<condition>\nCONDITION.\nINCORRECT: WHERE conditions are wrong.\n</condition>"),
]:
prompt = build_prompt(instr, schema, question, evidence, sql, es)
pairs.append({"prompt": prompt, "chosen": none_crit,
"rejected": fake_critique, # rejected = false alarm
"question": question, "db_id": d.get("db_id", ""),
"role": tag, "label": "correct"})
n_correct += 1
print(f" {src}: {n_wrong} wrong + {n_correct} correct examples")
rng.shuffle(pairs)
n_wrong_total = sum(1 for p in pairs if p["label"] == "wrong")
n_correct_total = sum(1 for p in pairs if p["label"] == "correct")
print(f"\nTotal ORPO pairs: {len(pairs)} ({n_wrong_total} wrong, {n_correct_total} correct)")
n_test = max(300, len(pairs) // 20)
test, train = pairs[:n_test], pairs[n_test:]
DatasetDict({
"train_dpo": Dataset.from_list(train),
"test_dpo": Dataset.from_list(test),
}).save_to_disk(OUT_DIR)
print(f"Saved {len(train)} train + {len(test)} test → {OUT_DIR}")
if __name__ == "__main__":
main()
|