File size: 10,583 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 | """
Validator v4 training data builder — augments v3 data with semantic-error negatives.
Problem: v3 validators have ~5.8% flagging rate on exec_ok=True wrong SQL.
Fix: add exec_ok=True wrong trajectories as negatives with heuristic critiques.
Output format: same as v3 (select/condition sections) for pipeline compatibility.
"""
import json, os, re, random, sqlite3, threading
from datasets import load_from_disk, 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",
]
V3_DATA = "data/sft-validator-selection-v3" # existing v3 data (sel)
V3_COND = "data/sft-validator-condition-v3" # existing v3 data (cond)
OUT_SEL = "data/hf_validator_v4_sel"
OUT_COND = "data/hf_validator_v4_cond"
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.")
def resolve_db_path(d):
db_path = d.get("db_path", "")
if db_path and os.path.exists(db_path):
return db_path
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_sql(db_path, sql, timeout=5):
result = [None]; error = [None]
def _run():
try:
conn = sqlite3.connect(db_path)
conn.text_factory = lambda b: b.decode(errors="ignore")
result[0] = conn.execute(sql).fetchmany(5)
conn.close()
except Exception as e:
error[0] = str(e)
t = threading.Thread(target=_run, daemon=True)
t.start(); t.join(timeout)
if t.is_alive():
return None, "TIMEOUT"
return result[0], error[0]
def generate_select_critique(wrong_sql, gold_sql):
"""Generate specific SELECT critique. Top errors from analysis:
DISTINCT mismatch (25.7%), aggregation mismatch (25.5%), subquery diff (12%)."""
wl, gl = wrong_sql.lower(), gold_sql.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()} in SELECT")
elif agg in wl and agg not in gl:
issues.append(f"Unexpected {agg[:-1].upper()} in SELECT")
if "distinct" in gl and "distinct" not in wl:
issues.append("Missing DISTINCT — query returns duplicate rows")
elif "distinct" in wl and "distinct" not in gl:
issues.append("Unexpected DISTINCT — query incorrectly deduplicates")
# Subquery difference
gs, ws = gl.count("select") - 1, wl.count("select") - 1
if gs > ws:
issues.append(f"Missing subquery (gold has {gs}, wrong has {ws})")
elif ws > gs:
issues.append(f"Unexpected subquery (gold has {gs}, wrong has {ws})")
if issues:
detail = "INCORRECT: " + "; ".join(issues) + "."
else:
detail = "INCORRECT: SELECT clause returns wrong results for this question."
return f"<select>\nSELECT.\n{detail}\n</select>"
def generate_condition_critique(wrong_sql, gold_sql):
"""Generate specific CONDITION critique. Top errors: JOIN mismatch (30%),
GROUP BY (6.9%), ORDER BY (8.3%), LIMIT (7.8%), subtle conditions (30.6%)."""
wl, gl = wrong_sql.lower(), gold_sql.lower()
issues = []
# JOIN count
gj, wj = gl.count("join"), wl.count("join")
if gj > wj:
issues.append(f"Missing JOIN (gold has {gj}, wrong has {wj})")
elif wj > gj:
issues.append(f"Extra JOIN (gold has {gj}, wrong has {wj})")
if "group by" in gl and "group by" not in wl:
issues.append("Missing GROUP BY clause")
elif "group by" in wl and "group by" not in gl:
issues.append("Unexpected GROUP BY clause")
if "having" in gl and "having" not in wl:
issues.append("Missing HAVING clause")
if ("order by" in gl) != ("order by" in wl):
issues.append("ORDER BY mismatch")
if ("limit" in gl) != ("limit" in wl):
issues.append("LIMIT clause mismatch")
if issues:
detail = "INCORRECT: " + "; ".join(issues) + "."
else:
detail = "INCORRECT: WHERE/HAVING conditions return wrong results for this question."
return f"<condition>\nCONDITION.\n{detail}\n</condition>"
NONE_SEL = "<select>\nSELECT.\nNone\n</select>"
NONE_COND = "<condition>\nCONDITION.\nNone\n</condition>"
def build_prompt(instr, schema, question, evidence, sql, exec_str):
# Field labels must match run_pipeline_rollouts.py VALIDATOR_PROMPT_BODY exactly.
body = (f"database schema:\n{schema}\n\nQuestion: {question}\n"
f"External knowledge: {evidence or 'None'}\n\nGenerated SQL query: {sql}\n\nExecution response:\n{exec_str}\n\n")
return instr + "\n\n" + body
def make_row(instr, schema, question, evidence, sql, exec_str, completion):
prompt = build_prompt(instr, schema, question, evidence, sql, exec_str)
# "chosen" key for train_fixer_v2.py compatibility; "completion" for legacy
return {"prompt": prompt, "chosen": completion, "completion": completion,
"messages": {"prompt": prompt, "completion": completion}}
def safe_trunc(s, n=3000):
s = str(s or "")
return s if len(s) <= n else s[:n] + "..."
def main():
rng = random.Random(42)
new_sel, new_cond = [], []
seen = set()
for src in SRC_PATHS:
if not os.path.exists(src):
print(f"skip {src}"); continue
n_pos = n_neg = 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_path(d)
if not db_path: continue
schema = safe_trunc(str(d.get("schema", "")), 2800)
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
correct = bool(t.get("is_planner_correct") or t.get("is_fixed_correct"))
exec_ok = bool(t.get("planner_exec_ok", True))
key = (hash(question), sql[:60])
if key in seen: continue
seen.add(key)
rows, err = exec_sql(db_path, sql)
if rows is not None:
exec_str = "OK. Rows: " + str(rows)[:300]
elif err:
exec_str = "Error: " + err[:200]
else:
exec_str = "No result"
if correct and exec_ok:
# POSITIVE: SQL is correct
r_sel = make_row(SEL_INSTR, schema, question, evidence, sql, exec_str, NONE_SEL)
r_cond = make_row(COND_INSTR, schema, question, evidence, sql, exec_str, NONE_COND)
new_sel.append(r_sel)
new_cond.append(r_cond)
n_pos += 1
elif not correct and exec_ok and gold_sql:
# NEGATIVE: exec_ok=True but wrong — semantic error
sel_crit = generate_select_critique(sql, gold_sql)
cond_crit = generate_condition_critique(sql, gold_sql)
r_sel = make_row(SEL_INSTR, schema, question, evidence, sql, exec_str, sel_crit)
r_cond = make_row(COND_INSTR, schema, question, evidence, sql, exec_str, cond_crit)
new_sel.append(r_sel)
new_cond.append(r_cond)
n_neg += 1
elif not exec_ok:
# NEGATIVE: exec error — clear signal
err_msg = exec_str[:200]
sel_crit = f"<select>\nSELECT.\nSQL fails to execute: {err_msg}\n</select>"
cond_crit = f"<condition>\nCONDITION.\nSQL fails to execute: {err_msg}\n</condition>"
r_sel = make_row(SEL_INSTR, schema, question, evidence, sql, exec_str, sel_crit)
r_cond = make_row(COND_INSTR, schema, question, evidence, sql, exec_str, cond_crit)
new_sel.append(r_sel)
new_cond.append(r_cond)
n_neg += 1
print(f" {src}: +{n_pos} pos, +{n_neg} neg")
print(f"\nNew rows — sel: {len(new_sel)}, cond: {len(new_cond)}")
# Load existing v3 data and merge
def load_v3(path):
d = load_from_disk(path)
rows = []
for split in ["train", "test", "train_dpo", "test_dpo"]:
if split in d:
for ex in d[split]:
p, c = ex["prompt"], ex["completion"]
rows.append({"prompt": p, "chosen": c, "completion": c,
"messages": {"prompt": p, "completion": c}})
return rows
v3_sel = load_v3(V3_DATA)
v3_cond = load_v3(V3_COND)
print(f"V3 existing — sel: {len(v3_sel)}, cond: {len(v3_cond)}")
combined_sel = v3_sel + new_sel
combined_cond = v3_cond + new_cond
rng.shuffle(combined_sel)
rng.shuffle(combined_cond)
def split_and_save(rows, out_dir):
n_test = max(200, len(rows) // 20)
test, train = rows[:n_test], rows[n_test:]
# Use train_dpo/test_dpo split names for train_fixer_v2.py compatibility
DatasetDict({
"train_dpo": Dataset.from_list(train),
"test_dpo": Dataset.from_list(test),
}).save_to_disk(out_dir)
print(f" saved {len(train)} train_dpo + {len(test)} test_dpo → {out_dir}")
split_and_save(combined_sel, OUT_SEL)
split_and_save(combined_cond, OUT_COND)
print("DONE")
if __name__ == "__main__":
main()
|