mats-sql-bundle / code /scripts /build_sft_griffith.py
thanhdath's picture
Push code: scripts, slurm sbatch, recipes, utils (v3 + selector series)
778d47d verified
Raw
History Blame Contribute Delete
6.8 kB
"""
Build SFT dataset from griffith-bigdata/sft_text2sql (deepseek-reasoner prompts)
with CoT completions generated by the thanhdath Llama-3B planner.
Pipeline:
1. Load griffith prompts (rich natural-language schema + External Knowledge + Question)
2. Reformat as our planning prompt (griffith schema + "Planning:" suffix)
3. Run thanhdath planner to generate CoT completions
4. Execute the predicted SQL and keep only correct predictions
5. Save as SFT dataset: (prompt, CoT_completion) pairs
Output: data/hf_planner_sft_griffith
Completion format: full CoT text ending with ```sql ... ``` (same as existing planner outputs)
"""
import json, os, re, sys, random, sqlite3, threading, requests
from datasets import Dataset, DatasetDict
ROOT = "/weka/s225250685/mats-tist"
os.chdir(ROOT)
sys.path.insert(0, ROOT)
from data_processing.planner import is_execution_correct
HF_CACHE = "/weka/s225250685/Huggingface/hub"
OUT = "data/hf_planner_sft_griffith"
PLANNER_URL = "http://localhost:8100" # thanhdath planner served externally before this script runs
PLANNER_MODEL = "planner"
MAX_TOKENS = 1024
TEMPERATURE = 0.0 # greedy for SFT data (deterministic, highest quality)
SEED = 42
print("Loading BIRD train gold SQL...", flush=True)
with open("data/sft_bird_with_evidence_train_text2sql.json") as f:
bird_train = json.load(f)
print(f"BIRD train: {len(bird_train)} questions", flush=True)
def safe_exec(db_path, sql, timeout=5):
result = [None]; err = [None]
def _run():
try:
conn = sqlite3.connect(db_path)
conn.text_factory = lambda b: b.decode(errors="ignore")
result[0] = conn.execute(sql).fetchmany(10)
conn.close()
except Exception as e:
err[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], err[0]
def extract_sql(cot_text):
"""Extract SQL from the planner CoT output (```sql...``` or ```...``` block)."""
m = re.search(r"```(?:sql)?\s*(.*?)\s*```", cot_text, re.DOTALL)
if m:
sql = m.group(1).strip()
if sql.upper().startswith("SQL"):
sql = sql[3:].strip()
return sql
# Fallback: last non-empty line
lines = [l.strip() for l in cot_text.strip().split("\n") if l.strip()]
return lines[-1] if lines else ""
def llama3_chat(prompt):
"""Build raw vLLM completion prompt in Llama-3 format."""
return (f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n"
f"{prompt}<|eot_id|>\n<|start_header_id|>assistant<|end_header_id|>\n")
def call_planner(prompt_text):
"""Call vLLM completions endpoint, return CoT text or None."""
raw_prompt = llama3_chat(prompt_text)
try:
r = requests.post(f"{PLANNER_URL}/v1/completions", json={
"model": PLANNER_MODEL,
"prompt": raw_prompt,
"max_tokens": MAX_TOKENS,
"temperature": TEMPERATURE,
"n": 1,
"seed": SEED,
"stop": ["<|eot_id|>"],
}, timeout=60)
r.raise_for_status()
return r.json()["choices"][0]["text"].strip()
except Exception as e:
return None
# Build prompt: griffith schema text + our "Planning:" suffix
# The griffith user message already has:
# "Database Schema:\n...\nExternal Knowledge:\n...\nQuestion: ...\n"
# We append "Planning:" to trigger CoT output.
def build_planning_prompt(griffith_user_msg):
"""Convert griffith user message to our planning prompt format."""
# Strip trailing whitespace, then append Planning: trigger
return griffith_user_msg.rstrip() + "\n\nPlanning:"
print("Loading griffith-bigdata/sft_text2sql (deepseek-reasoner)...", flush=True)
from datasets import load_dataset
ds_raw = load_dataset("griffith-bigdata/sft_text2sql", split="train_sft", cache_dir=HF_CACHE)
ds_dr = ds_raw.filter(lambda x: x["model_name"] == "deepseek-reasoner")
print(f"deepseek-reasoner rows: {len(ds_dr)}", flush=True)
rows = []
n_correct = 0
n_wrong = 0
n_skip = 0
n_qmismatch = 0
for i, row in enumerate(ds_dr):
sid = int(row["sample_id"])
if sid < 0 or sid >= len(bird_train):
n_skip += 1
continue
msgs = row["messages"]
user_msg = msgs[1]["content"] # griffith schema + evidence + question
# Cross-check question matches BIRD
q_match = re.search(r"Question:\s*(.+?)(?:\n|$)", user_msg)
griffith_q = q_match.group(1).strip() if q_match else ""
bird_q = bird_train[sid]["question"].strip()
if griffith_q.lower() != bird_q.lower():
n_qmismatch += 1
n_skip += 1
continue
gold_sql = bird_train[sid]["sql"]
db_id = bird_train[sid].get("db_id", "")
db_path = (bird_train[sid].get("db_path") or
f"data/train_databases/{db_id}/{db_id}.sqlite")
# Build planning prompt
planning_prompt = build_planning_prompt(user_msg)
# Run thanhdath planner (greedy)
cot_text = call_planner(planning_prompt)
if not cot_text:
n_wrong += 1
continue
# Extract and execute predicted SQL
pred_sql = extract_sql(cot_text)
if not pred_sql:
n_wrong += 1
continue
gold_res, _ = safe_exec(db_path, gold_sql)
pred_res, err = safe_exec(db_path, pred_sql)
if err or not is_execution_correct(gold_res, pred_res):
n_wrong += 1
continue
# Correct prediction — keep this (prompt, CoT) pair
rows.append({
"prompt": planning_prompt, # griffith schema + Planning: trigger
"completion": cot_text, # full CoT: Goal→Condition→Tables→SQL
"sample_id": sid,
"db_id": db_id,
"question": bird_q,
"gold_sql": gold_sql,
})
n_correct += 1
if (i + 1) % 500 == 0:
print(f" [{i+1}/{len(ds_dr)}] correct={n_correct} wrong={n_wrong} skip={n_skip}", flush=True)
print(f"\nFinal: {n_correct} correct / {n_correct+n_wrong} attempted ({n_correct/(n_correct+n_wrong)*100:.1f}% pass rate)", flush=True)
print(f"Skipped: {n_skip} (q_mismatch={n_qmismatch})", flush=True)
if not rows:
print("ERROR: no correct pairs collected — check planner endpoint", flush=True)
sys.exit(1)
# Sanity check: show 3 examples
for ex in rows[:3]:
print(f"\n sid={ex['sample_id']} db={ex['db_id']}", flush=True)
print(f" Q: {ex['question'][:70]}", flush=True)
print(f" CoT: {ex['completion'][:120]}...", flush=True)
# 90/10 split
random.seed(42)
random.shuffle(rows)
n_train = int(0.9 * len(rows))
DatasetDict({
"train": Dataset.from_list(rows[:n_train]),
"test": Dataset.from_list(rows[n_train:]),
}).save_to_disk(OUT)
print(f"\nSaved → {OUT} (train={n_train}, test={len(rows)-n_train})", flush=True)