Spaces:
Running
Running
File size: 11,121 Bytes
a39d8ef | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 | """
generate_edge_cases.py
======================
Targeted edge-case data generator for the 4 failure patterns found in eval:
1. ROW_NUMBER vs RANK vs DENSE_RANK (tie-breaking semantics)
2. strftime month as INTEGER (not '%Y-%m' string)
3. SELECT column discipline (no unrequested extras)
4. LAG/LEAD period-over-period
5. HAVING vs WHERE placement
6. COUNT(DISTINCT) vs COUNT
Produces: edge_cases.jsonl (same chat format as nl2sql_cleaned_ready_to_train.jsonl)
Run: python generate_edge_cases.py
"""
import os, sys, json, re, hashlib
from tqdm import tqdm
from transformers import AutoModelForCausalLM, AutoTokenizer, AwqConfig, BitsAndBytesConfig
import torch
import transformers.activations
# Yeh line AutoAWQ ko bewakoof banayegi taaki wo crash na ho
if not hasattr(transformers.activations, 'PytorchGELUTanh'):
transformers.activations.PytorchGELUTanh = transformers.activations.NewGELUActivation
os.environ["CUDA_VISIBLE_DEVICES"] = "3,1,6,7"
PROJECT_ROOT = os.path.abspath(os.path.dirname(__file__))
if PROJECT_ROOT not in sys.path:
sys.path.insert(0, PROJECT_ROOT)
from data_factory.schemas import SCHEMA_CONTEXT
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4"
)
MODEL_NAME = "Qwen/Qwen2.5-72B-Instruct"
OUTPUT_FILE = "edge_cases.jsonl"
BATCH_SIZE = 8 # smaller β edge prompts are long
SAMPLES_PER_PATTERN = 715 # ~6 batches per pattern β 5005 total edge samples
SYSTEM_PROMPT = (
"You are a Senior SQL Architect. "
"Output ONLY the SQL query. Use SQLite syntax."
)
# ββ Edge-case prompt templates ββββββββββββββββββββββββββββββββββββββββββββββ
# Each entry: (pattern_tag, user_prompt_template)
# {schema} is filled at runtime with a random domain schema.
EDGE_PATTERNS = [
# 1. ROW_NUMBER tie-breaking β the #1 failure
("row_number_tiebreak", """SCHEMA:
{schema}
Generate exactly 8 NL2SQL pairs that REQUIRE ROW_NUMBER() (not RANK or DENSE_RANK) \
because the question explicitly says "pick one winner when there is a tie" \
using a tiebreaker column (e.g. lower id, earlier date).
Output ONLY a valid JSON array:
[
{{"nl": "...", "sql": "SELECT ..."}},
...
]
Rules:
- Every SQL must use ROW_NUMBER() OVER (...) not RANK().
- The OVER clause ORDER BY must include the tiebreaker column.
- WHERE rn = 1 must appear in an outer query or CTE.
- No markdown. No explanation. Just the JSON array."""),
# 2. RANK / DENSE_RANK β when ties SHOULD persist
("rank_dense_rank", """SCHEMA:
{schema}
Generate exactly 8 NL2SQL pairs where RANK() or DENSE_RANK() is the CORRECT choice \
because the question says "show all tied records at the same rank".
Output ONLY a valid JSON array:
[
{{"nl": "...", "sql": "SELECT ..."}},
...
]
Rules:
- Use RANK() when question implies gaps after ties, DENSE_RANK() when no gaps.
- NL must make the tie-semantics explicit ("same rank", "tied positions").
- No markdown. No explanation. Just the JSON array."""),
# 3. strftime integer month output
("strftime_integer_month", """SCHEMA:
{schema}
Generate exactly 8 NL2SQL pairs where the question asks for a numeric month number \
(1β12), NOT a 'YYYY-MM' string.
Output ONLY a valid JSON array:
[
{{"nl": "...", "sql": "SELECT ..."}},
...
]
Rules:
- SQL must use CAST(strftime('%m', <col>) AS INTEGER) to produce integer month.
- Do NOT use strftime('%Y-%m', ...) when the question asks for month number.
- NL questions must say "month number", "which month (1β12)", or similar.
- No markdown. No explanation. Just the JSON array."""),
# 4. SELECT column discipline
("select_column_discipline", """SCHEMA:
{schema}
Generate exactly 8 NL2SQL pairs where the question explicitly names ONLY the columns \
to return. The SQL must select EXACTLY those columns β no extras like avg_salary, \
row counts, or intermediate aggregates.
Output ONLY a valid JSON array:
[
{{"nl": "...", "sql": "SELECT ..."}},
...
]
Rules:
- NL must say "return only X, Y, Z" or "show me only the name and total".
- SQL SELECT list must contain only those columns.
- If aggregation is needed internally (e.g. for HAVING), do NOT expose it in SELECT.
- No markdown. No explanation. Just the JSON array."""),
# 5. LAG / LEAD period-over-period
("lag_lead_period", """SCHEMA:
{schema}
Generate exactly 8 NL2SQL pairs that require LAG() or LEAD() window functions \
for period-over-period comparison (e.g. month-over-month revenue change, \
previous order amount, next appointment date).
Output ONLY a valid JSON array:
[
{{"nl": "...", "sql": "SELECT ..."}},
...
]
Rules:
- Use LAG(<col>, 1) OVER (ORDER BY ...) or LEAD(...) correctly.
- NL must imply comparison with previous or next row/period.
- No markdown. No explanation. Just the JSON array."""),
# 6. HAVING vs WHERE
("having_vs_where", """SCHEMA:
{schema}
Generate exactly 8 NL2SQL pairs that test correct placement of filter conditions:
- Conditions on raw columns β WHERE
- Conditions on aggregates β HAVING
Include 4 pairs where a wrong model might put an aggregate condition in WHERE (trap).
Output ONLY a valid JSON array:
[
{{"nl": "...", "sql": "SELECT ..."}},
...
]
Rules:
- SQL must never filter an aggregate (COUNT, SUM, AVG) inside WHERE.
- SQL must never put a raw column filter inside HAVING.
- No markdown. No explanation. Just the JSON array."""),
# 7. COUNT(DISTINCT) vs COUNT
("count_distinct", """SCHEMA:
{schema}
Generate exactly 8 NL2SQL pairs where the question specifically asks for \
"unique", "distinct", or "different" counts β requiring COUNT(DISTINCT col).
Also include 2 pairs where COUNT(*) is correct to reinforce the contrast.
Output ONLY a valid JSON array:
[
{{"nl": "...", "sql": "SELECT ..."}},
...
]
Rules:
- When NL says "unique/distinct", SQL must use COUNT(DISTINCT <col>).
- When NL says "total orders placed" (not distinct), use COUNT(*) or COUNT(id).
- No markdown. No explanation. Just the JSON array."""),
]
# ββ Helpers ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def extract_json_array(text: str) -> str:
text = text.strip()
# strip code fences if model leaks them
text = re.sub(r"```(?:json)?\n?(.*?)```", r"\1", text, flags=re.DOTALL).strip()
s, e = text.find("["), text.rfind("]")
return text[s:e+1] if s != -1 and e != -1 else "[]"
def get_hash(text: str) -> str:
return hashlib.md5(text.lower().strip().encode()).hexdigest()
def build_record(nl: str, sql: str, domain: str) -> dict:
return {
"prompt": [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": f"SCHEMA: {SCHEMA_CONTEXT[domain]}\nQUESTION: {nl}"}
],
"sql": sql
}
# ββ Main βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def main():
print(f"Loading {MODEL_NAME}...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, padding_side="left")
tokenizer.pad_token = tokenizer.eos_token
custom_memory = {0:"30GiB",1:"75GiB",2:"45GiB",3:"45GiB"}
model = AutoModelForCausalLM.from_pretrained(
MODEL_NAME,
device_map="auto",
max_memory=custom_memory,
quantization_config=quantization_config,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
attn_implementation = "sdpa"
)
domains = list(SCHEMA_CONTEXT.keys())
seen = set()
total = 0
out = open(OUTPUT_FILE, "a", encoding="utf-8")
for pattern_tag, prompt_tmpl in EDGE_PATTERNS:
print(f"\n[PATTERN] {pattern_tag}")
collected = 0
domain_idx = 0
pbar = tqdm(total=SAMPLES_PER_PATTERN, desc=pattern_tag)
while collected < SAMPLES_PER_PATTERN:
# Build a batch of prompts, cycling through domains
batch_domains = []
batch_prompts = []
for _ in range(BATCH_SIZE):
domain = domains[domain_idx % len(domains)]
domain_idx += 1
user_msg = prompt_tmpl.format(schema=SCHEMA_CONTEXT[domain])
msgs = [
{"role": "system", "content": "You output only valid JSON arrays. No markdown."},
{"role": "user", "content": user_msg}
]
batch_prompts.append(
tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
)
batch_domains.append(domain)
inputs = tokenizer(
batch_prompts, return_tensors="pt", padding=True, truncation=True
).to(model.device)
try:
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=2048,
do_sample=True,
temperature=0.5,
top_p=0.9,
pad_token_id=tokenizer.eos_token_id
)
responses = tokenizer.batch_decode(
outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True
)
for resp, domain in zip(responses, batch_domains):
raw = extract_json_array(resp)
try:
pairs = json.loads(raw)
except Exception:
continue
for pair in pairs:
nl = pair.get("nl", "").strip()
sql = pair.get("sql", "").strip()
if not nl or not sql:
continue
# strip fences in sql just in case
sql = re.sub(r"```(?:sql)?\n?(.*?)```", r"\1", sql, flags=re.DOTALL).strip()
h = get_hash(nl + sql)
if h in seen:
continue
seen.add(h)
record = build_record(nl, sql, domain)
out.write(json.dumps(record, ensure_ascii=False) + "\n")
out.flush()
collected += 1
total += 1
pbar.update(1)
if collected >= SAMPLES_PER_PATTERN:
break
except Exception as e:
tqdm.write(f"[WARN] Batch failed: {e}")
continue
pbar.close()
out.close()
print(f"\nDone! {total} edge-case records saved to {OUTPUT_FILE}")
if __name__ == "__main__":
main() |