Spaces:
Running
Running
Commit ·
051d18c
1
Parent(s): 61b8743
feat(round2): build Stage 1 Data Repair task and string-based deterministic grader
Browse files- app/graders/grader_stage1.py +69 -0
- app/tasks/stage1_data_repair.py +98 -0
app/graders/grader_stage1.py
ADDED
|
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from __future__ import annotations
|
| 2 |
+
|
| 3 |
+
import re
|
| 4 |
+
from typing import Optional
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
def _contains_accuracy(output: str) -> bool:
|
| 8 |
+
"""Check if output contains a printed accuracy value."""
|
| 9 |
+
return bool(re.search(r"Accuracy\s*:\s*\d*\.?\d+", output, flags=re.IGNORECASE))
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
def _has_nan_handling(script: str) -> bool:
|
| 13 |
+
"""Detect common NaN/missing-value handling patterns in script."""
|
| 14 |
+
s = script.lower()
|
| 15 |
+
return any(
|
| 16 |
+
k in s for k in ["dropna", "fillna", "simpleimputer", "isna()", "notna()"]
|
| 17 |
+
)
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def _has_outlier_handling(script: str) -> bool:
|
| 21 |
+
"""Detect common outlier handling patterns in script."""
|
| 22 |
+
s = script.lower()
|
| 23 |
+
return any(
|
| 24 |
+
k in s for k in ["clip(", "quantile(", "iqr", "winsor", "robustscaler"]
|
| 25 |
+
)
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
def grade(script: str, run_output: str, run_error: Optional[str]) -> float:
|
| 29 |
+
"""
|
| 30 |
+
Stage 1 Grader — Data Repair.
|
| 31 |
+
|
| 32 |
+
Scoring ladder (deterministic, string-based only, no script execution):
|
| 33 |
+
0.0 → KeyError still present (age_years in script)
|
| 34 |
+
0.2 → age_years fixed but NaN error still occurs
|
| 35 |
+
0.4 → script contains NaN handling (dropna/fillna/etc.)
|
| 36 |
+
0.6 → script contains outlier handling (clip/quantile/etc.)
|
| 37 |
+
0.8 → script runs without any error
|
| 38 |
+
1.0 → script runs clean AND prints accuracy value
|
| 39 |
+
|
| 40 |
+
Checks from top (1.0) downward, returns first match.
|
| 41 |
+
"""
|
| 42 |
+
run_error = run_error or ""
|
| 43 |
+
run_output = run_output or ""
|
| 44 |
+
script = script or ""
|
| 45 |
+
|
| 46 |
+
# ── Hard failure: column rename bug still present ──
|
| 47 |
+
if "age_years" in script:
|
| 48 |
+
return 0.0
|
| 49 |
+
|
| 50 |
+
# ── Top tier: clean run with accuracy printed ──
|
| 51 |
+
if run_error.strip() == "" and _contains_accuracy(run_output):
|
| 52 |
+
return 1.0
|
| 53 |
+
|
| 54 |
+
# ── Clean run but no accuracy printed ──
|
| 55 |
+
if run_error.strip() == "":
|
| 56 |
+
return 0.8
|
| 57 |
+
|
| 58 |
+
# ── Script still has errors, but check for fix progress ──
|
| 59 |
+
|
| 60 |
+
# Outlier handling present (implies NaN handling also likely attempted)
|
| 61 |
+
if _has_outlier_handling(script) and _has_nan_handling(script):
|
| 62 |
+
return 0.6
|
| 63 |
+
|
| 64 |
+
# NaN handling present
|
| 65 |
+
if _has_nan_handling(script):
|
| 66 |
+
return 0.4
|
| 67 |
+
|
| 68 |
+
# Column rename fixed but script still crashes (NaN error)
|
| 69 |
+
return 0.2
|
app/tasks/stage1_data_repair.py
ADDED
|
@@ -0,0 +1,98 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import numpy as np
|
| 2 |
+
import pandas as pd
|
| 3 |
+
|
| 4 |
+
|
| 5 |
+
def generate_scenario() -> dict:
|
| 6 |
+
"""
|
| 7 |
+
Stage 1 — Data Repair (combines old Task 1 + Task 2).
|
| 8 |
+
|
| 9 |
+
Two-bug cascade:
|
| 10 |
+
Bug 1: Script references 'age_years' but DataFrame column is 'age' → KeyError
|
| 11 |
+
Bug 2: NaN values in data cause StandardScaler to crash → ValueError
|
| 12 |
+
|
| 13 |
+
The agent must make TWO edit_script actions to fix both bugs sequentially.
|
| 14 |
+
|
| 15 |
+
Data: 500 rows, seed=42, with ~15% NaN injection and 10 salary outliers.
|
| 16 |
+
"""
|
| 17 |
+
|
| 18 |
+
rng = np.random.default_rng(42)
|
| 19 |
+
n = 500
|
| 20 |
+
|
| 21 |
+
age = rng.integers(18, 71, size=n).astype(float)
|
| 22 |
+
salary = np.clip(rng.normal(70000, 20000, size=n), 20000, None)
|
| 23 |
+
credit_score = np.clip(rng.normal(680, 80, size=n), 300, 850)
|
| 24 |
+
loan_amount = np.clip(rng.normal(20000, 10000, size=n), 1000, None)
|
| 25 |
+
employment_years = np.clip(rng.normal(8, 5, size=n), 0, 40).astype(float)
|
| 26 |
+
|
| 27 |
+
# Binary target influenced by features with noise
|
| 28 |
+
logits = (
|
| 29 |
+
(age - 40) * -0.03
|
| 30 |
+
+ (salary - 70000) / 50000
|
| 31 |
+
+ (credit_score - 650) * 0.004
|
| 32 |
+
- (loan_amount - 20000) / 40000
|
| 33 |
+
+ (employment_years - 8) * 0.05
|
| 34 |
+
+ rng.normal(0, 0.8, size=n)
|
| 35 |
+
)
|
| 36 |
+
prob = 1 / (1 + np.exp(-logits))
|
| 37 |
+
target = (prob > 0.5).astype(int)
|
| 38 |
+
|
| 39 |
+
df = pd.DataFrame(
|
| 40 |
+
{
|
| 41 |
+
"age": age,
|
| 42 |
+
"salary": salary,
|
| 43 |
+
"credit_score": credit_score,
|
| 44 |
+
"loan_amount": loan_amount,
|
| 45 |
+
"employment_years": employment_years,
|
| 46 |
+
"target": target,
|
| 47 |
+
}
|
| 48 |
+
)
|
| 49 |
+
|
| 50 |
+
# Inject ~15% rows with NaNs across a subset of columns
|
| 51 |
+
nan_rows = rng.choice(np.arange(n), size=int(0.15 * n), replace=False)
|
| 52 |
+
for col in ["age", "credit_score", "employment_years"]:
|
| 53 |
+
df.loc[nan_rows, col] = np.nan
|
| 54 |
+
|
| 55 |
+
# Inject ~10 extreme salary outliers (non-overlapping with NaN rows)
|
| 56 |
+
outlier_rows = rng.choice(
|
| 57 |
+
np.setdiff1d(np.arange(n), nan_rows), size=10, replace=False
|
| 58 |
+
)
|
| 59 |
+
df.loc[outlier_rows, "salary"] = 9_999_999.0
|
| 60 |
+
|
| 61 |
+
broken_script = (
|
| 62 |
+
"import pandas as pd\n"
|
| 63 |
+
"from sklearn.preprocessing import StandardScaler\n"
|
| 64 |
+
"from sklearn.linear_model import LogisticRegression\n"
|
| 65 |
+
"from sklearn.model_selection import train_test_split\n"
|
| 66 |
+
"\n"
|
| 67 |
+
"X = df[['age_years', 'salary', 'credit_score', 'loan_amount', 'employment_years']].copy()\n"
|
| 68 |
+
"y = df['target']\n"
|
| 69 |
+
"scaler = StandardScaler()\n"
|
| 70 |
+
"X_scaled = scaler.fit_transform(X)\n"
|
| 71 |
+
"X_train, X_test, y_train, y_test = train_test_split(X_scaled, y, test_size=0.2, random_state=42)\n"
|
| 72 |
+
"from sklearn.linear_model import LogisticRegression\n"
|
| 73 |
+
"clf = LogisticRegression(max_iter=1000, random_state=42)\n"
|
| 74 |
+
"clf.fit(X_train, y_train)\n"
|
| 75 |
+
"print('Accuracy:', clf.score(X_test, y_test))\n"
|
| 76 |
+
)
|
| 77 |
+
|
| 78 |
+
initial_error_log = (
|
| 79 |
+
"Traceback (most recent call last):\n"
|
| 80 |
+
" File '<script>', line 6, in <module>\n"
|
| 81 |
+
"KeyError: 'age_years'\n"
|
| 82 |
+
)
|
| 83 |
+
|
| 84 |
+
task_description = (
|
| 85 |
+
"Fix the broken pipeline: there are multiple bugs. "
|
| 86 |
+
"Find and fix all of them."
|
| 87 |
+
)
|
| 88 |
+
|
| 89 |
+
expected_fix_hint = "fix age_years→age AND add dropna/clip before scaler"
|
| 90 |
+
|
| 91 |
+
return {
|
| 92 |
+
"broken_script": broken_script,
|
| 93 |
+
"dataframe": df,
|
| 94 |
+
"initial_error_log": initial_error_log,
|
| 95 |
+
"task_description": task_description,
|
| 96 |
+
"expected_fix_hint": expected_fix_hint,
|
| 97 |
+
"stage_number": 1,
|
| 98 |
+
}
|