Data-cleaning / environment.py
Mihir Mungara
graders result clampped between 0 and 1
747aba3
Raw
History Blame Contribute Delete
23.3 kB
import json
import pandas as pd
import numpy as np
from typing import Any, Dict, Optional, Tuple
from models import Action, Observation, Reward, StepResult, TaskInfo
AVAILABLE_OPERATIONS = [
"remove_duplicates",
"fill_missing_mean",
"fill_missing_mode",
"fill_missing_median",
"fix_dtype",
"remove_outliers",
"rename_columns",
"validate_schema",
"finish"
]
class DataCleaningEnv:
"""
OpenEnv-compliant Data Cleaning Environment.
Agent must clean dirty datasets step by step.
"""
def __init__(self, task_id: str = "easy_dedup_rename"):
self.task_id = task_id
self.current_df: Optional[pd.DataFrame] = None
self.gold_df: Optional[pd.DataFrame] = None
self.task_meta: Dict = {}
self.step_count: int = 0
self.done: bool = False
self.max_steps: int = 10
self.reward_history = []
self.actions_taken: List[str] = [] # Track sequence of actions
self._load_task_metadata()
# ─────────────────────────────────────────
# INTERNAL HELPERS
# ─────────────────────────────────────────
def _load_task_metadata(self):
with open("datasets/task_metadata.json", "r") as f:
all_meta = json.load(f)
mapping = {
"easy_dedup_rename": "easy",
"medium_missing_dtype": "medium",
"hard_full_pipeline": "hard",
"expert_sales_pipeline": "expert"
}
key = mapping.get(self.task_id, "easy")
self.task_meta = all_meta[key]
self.max_steps = self.task_meta["max_steps"]
def _load_datasets(self):
mapping = {
"easy_dedup_rename": "easy",
"medium_missing_dtype": "medium",
"hard_full_pipeline": "hard",
"expert_sales_pipeline": "expert"
}
folder = mapping.get(self.task_id, "easy")
self.current_df = pd.read_csv(f"datasets/{folder}/dirty.csv")
self.gold_df = pd.read_csv(f"datasets/{folder}/gold.csv")
def _add_random_variation(self, df: pd.DataFrame) -> pd.DataFrame:
"""Add small random variations to prevent memorization."""
df = df.copy()
rng = np.random.default_rng()
for col in df.columns:
# Randomly shuffle some numeric values slightly
num_series = pd.to_numeric(df[col], errors="coerce")
if num_series.notna().sum() > 2:
# Add tiny noise to 1-2 numeric cells
idx = df.index[num_series.notna()].tolist()
if len(idx) >= 2:
pick = rng.choice(idx, size=min(2, len(idx)), replace=False)
for i in pick:
orig = num_series[i]
noise = rng.integers(-3, 4)
df.at[i, col] = str(int(orig + noise)) if isinstance(df.at[i, col], str) else orig + noise
# Randomly shuffle row order (keeps same data, different order)
df = df.sample(frac=1, random_state=rng.integers(0, 9999)).reset_index(drop=True)
return df
def _get_observation(self, message: str = "") -> Observation:
df = self.current_df
missing = {col: int(df[col].isnull().sum()) for col in df.columns}
dtypes = {col: str(df[col].dtype) for col in df.columns}
sample = df.head(3).fillna("NULL").to_dict(orient="records")
return Observation(
task_id=self.task_id,
step=self.step_count,
dataset_info={
"total_rows": len(df),
"total_columns": len(df.columns),
"has_duplicates": bool(df.duplicated().any()),
"has_missing": bool(df.isnull().any().any()),
},
columns=list(df.columns),
shape=[len(df), len(df.columns)],
missing_values=missing,
dtypes=dtypes,
duplicate_count=int(df.duplicated().sum()),
sample_rows=sample,
available_operations=self.task_meta.get(
"operations_allowed", AVAILABLE_OPERATIONS
),
task_description=self.task_meta.get("description", ""),
message=message
)
def _compute_sequence_penalty(self) -> float:
"""
Penalize illogical action sequences.
Optimal sequence:
1. remove_duplicates (clean up redundant data)
2. fix_dtype (understand structure)
3. fill_missing (based on correct types)
4. remove_outliers (after understanding distribution)
5. validate_schema (final check)
Penalties:
- Doing operations out of order: -0.05 per violation
- Repeating same operation: -0.02 per repeat
"""
OPTIMAL_ORDER = [
"remove_duplicates",
"fix_dtype",
"fill_missing_mean",
"fill_missing_mode",
"fill_missing_median",
"remove_outliers",
"validate_schema",
]
penalty = 0.0
last_order_idx = -1
prev_action = None
# Track positions of operations in optimal order
for action in self.actions_taken:
if action == "finish":
continue
# Penalize repeated same actions (doing same thing twice in a row)
if action == prev_action and action not in ["fill_missing_mean", "fill_missing_mode", "fill_missing_median", "remove_outliers"]:
penalty += 0.02
# Find this action in optimal order
if action in OPTIMAL_ORDER:
current_idx = OPTIMAL_ORDER.index(action)
# Penalize out-of-order operations
if current_idx < last_order_idx:
# Going backward in sequence (e.g., doing remove_duplicates after fill_missing)
penalty += 0.08
last_order_idx = current_idx
else:
# Unknown operation
penalty += 0.01
prev_action = action
return min(0.25, penalty) # Cap penalty at 0.25
def _compute_reward(self) -> Reward:
df = self.current_df
gold = self.gold_df
scoring = self.task_meta.get("scoring", {})
dup_score = 0.0
missing_score = 0.0
dtype_score = 0.0
outlier_score = 0.0
schema_score = 0.0
penalty = 0.0
# ── Duplicate score ──────────────────────────────────────────
if "duplicate_score" in scoring:
gold_rows = len(gold)
curr_rows = len(df)
if curr_rows == gold_rows:
dup_score = 1.0
elif curr_rows < gold_rows:
dup_score = max(0.0, curr_rows / gold_rows)
else:
excess = curr_rows - gold_rows
dup_score = max(0.0, 1.0 - (excess / gold_rows))
# ── Missing value score ──────────────────────────────────────
if "missing_score" in scoring:
# Count original missing values from dirty dataset
import os
mapping = {
"easy_dedup_rename": "easy",
"medium_missing_dtype": "medium",
"hard_full_pipeline": "hard",
"expert_sales_pipeline": "expert"
}
folder = mapping.get(self.task_id, "easy")
orig_df = pd.read_csv(f"datasets/{folder}/dirty.csv")
orig_missing = int(orig_df.isnull().sum().sum())
curr_missing = int(df.isnull().sum().sum())
gold_missing = int(gold.isnull().sum().sum())
if orig_missing <= gold_missing:
missing_score = 1.0
else:
filled_needed = orig_missing - gold_missing
filled_done = orig_missing - curr_missing
filled_done = max(0, filled_done)
missing_score = filled_done / filled_needed
missing_score = max(0.0, min(1.0, missing_score))
# ── Dtype score ──────────────────────────────────────────────
if "dtype_score" in scoring:
common_cols = [c for c in gold.columns if c in df.columns]
if common_cols:
matches = sum(
1 for c in common_cols
if str(df[c].dtype) == str(gold[c].dtype)
)
dtype_score = matches / len(common_cols)
# ── Outlier score ────────────────────────────────────────────
if "outlier_score" in scoring:
num_cols = gold.select_dtypes(include=[np.number]).columns
scores = []
for col in num_cols:
if col not in df.columns:
continue
gold_mean = gold[col].mean()
gold_std = gold[col].std() + 1e-9
curr_col = pd.to_numeric(df[col], errors="coerce").dropna()
outliers = ((curr_col - gold_mean).abs() > 3 * gold_std).sum()
col_score = max(0.0, 1.0 - outliers / (len(curr_col) + 1e-9))
scores.append(col_score)
outlier_score = float(np.mean(scores)) if scores else 0.0
# ── Schema score ─────────────────────────────────────────────
if "schema_score" in scoring:
gold_cols = list(gold.columns)
curr_cols = list(df.columns)
matched = sum(1 for c in gold_cols if c in curr_cols)
schema_score = matched / len(gold_cols) if gold_cols else 0.0
# ── Penalty for too many steps + sequence violations ─────────────
step_ratio = self.step_count / self.max_steps
base_penalty = 0.0
if step_ratio > 0.8:
base_penalty = 0.05
# Add sequence-based penalty
sequence_penalty = self._compute_sequence_penalty()
penalty = base_penalty + sequence_penalty
# ── Weighted total ───────────────────────────────────────────
weights = {
"duplicate_score": scoring.get("duplicate_score", 0.0),
"missing_score": scoring.get("missing_score", 0.0),
"dtype_score": scoring.get("dtype_score", 0.0),
"outlier_score": scoring.get("outlier_score", 0.0),
"schema_score": scoring.get("schema_score", 0.0),
}
component_scores = {
"duplicate_score": dup_score,
"missing_score": missing_score,
"dtype_score": dtype_score,
"outlier_score": outlier_score,
"schema_score": schema_score,
}
total = sum(
component_scores[k] * w
for k, w in weights.items()
) - penalty
# Clamp ALL scores strictly to (0, 1) β€” grader rejects 0.0 and 1.0
def _sc(v):
return round(max(0.0001, min(0.9999, float(v))), 4)
return Reward(
total=_sc(total),
duplicate_score=_sc(dup_score),
missing_score=_sc(missing_score),
dtype_score=_sc(dtype_score),
outlier_score=_sc(outlier_score),
schema_score=_sc(schema_score),
penalty=round(max(0.0, min(0.9999, float(penalty))), 4)
)
# ─────────────────────────────────────────
# OPERATIONS
# ─────────────────────────────────────────
def _op_remove_duplicates(self, params: Dict) -> str:
before = len(self.current_df)
subset = params.get("subset", None)
self.current_df = self.current_df.drop_duplicates(subset=subset)
self.current_df = self.current_df.reset_index(drop=True)
removed = before - len(self.current_df)
return f"Removed {removed} duplicate rows. Rows: {before} -> {len(self.current_df)}"
def _op_fill_missing(self, params: Dict) -> str:
col = params.get("column")
strategy = params.get("strategy", "mean")
messages = []
cols_to_fill = [col] if col else list(self.current_df.columns)
for c in cols_to_fill:
if self.current_df[c].isnull().sum() == 0:
continue
# For mean/median strategies, only apply to columns that are already numeric or can be safely converted
if strategy in ["mean", "median"]:
# Check if column is already numeric
if pd.api.types.is_numeric_dtype(self.current_df[c]):
# Column is already numeric, apply strategy directly
if strategy == "mean":
fill_val = self.current_df[c].mean()
self.current_df[c] = self.current_df[c].fillna(round(fill_val, 2))
elif strategy == "median":
fill_val = self.current_df[c].median()
self.current_df[c] = self.current_df[c].fillna(fill_val)
messages.append(f"{c}->{strategy}")
else:
# Try to convert to numeric - if it works for most values, use it
numeric = pd.to_numeric(self.current_df[c], errors="coerce")
# Only apply if at least 50% of non-null values are numeric
non_null_count = self.current_df[c].notna().sum()
numeric_count = numeric.notna().sum()
if non_null_count > 0 and numeric_count / non_null_count >= 0.5:
if strategy == "mean":
fill_val = numeric.mean()
self.current_df[c] = numeric.fillna(round(fill_val, 2))
elif strategy == "median":
fill_val = numeric.median()
self.current_df[c] = numeric.fillna(fill_val)
messages.append(f"{c}->{strategy}")
# else: skip this column as it's not suitable for numeric strategy
elif strategy == "mode":
mode_vals = self.current_df[c].mode()
if len(mode_vals) == 0:
continue
fill_val = mode_vals.iloc[0]
self.current_df[c] = self.current_df[c].fillna(fill_val)
messages.append(f"{c}->{strategy}")
elif strategy == "ffill":
self.current_df[c] = self.current_df[c].ffill()
messages.append(f"{c}->ffill")
else:
# Custom fill value
fill_val = strategy
self.current_df[c] = self.current_df[c].fillna(fill_val)
messages.append(f"{c}->{strategy}")
return f"Filled missing values: {', '.join(messages)}" if messages else "No missing values to fill"
def _op_fix_dtype(self, params: Dict) -> str:
col = params.get("column")
dtype = params.get("dtype", "auto")
messages = []
cols_to_fix = [col] if col else list(self.current_df.columns)
for c in cols_to_fix:
try:
if dtype == "int" or dtype == "auto":
converted = pd.to_numeric(self.current_df[c], errors="coerce")
if converted.notna().all():
self.current_df[c] = converted.astype(int)
messages.append(f"{c}->int")
elif dtype == "float":
self.current_df[c] = pd.to_numeric(
self.current_df[c], errors="coerce"
)
messages.append(f"{c}->float")
elif dtype == "str":
self.current_df[c] = self.current_df[c].astype(str)
messages.append(f"{c}->str")
except Exception as e:
messages.append(f"{c}->failed({e})")
return f"Fixed dtypes: {', '.join(messages)}"
def _op_remove_outliers(self, params: Dict) -> str:
col = params.get("column")
method = params.get("method", "iqr")
before = len(self.current_df)
messages = []
cols = [col] if col else list(
self.current_df.select_dtypes(include=[np.number]).columns
)
for c in cols:
series = pd.to_numeric(self.current_df[c], errors="coerce")
if method == "iqr":
Q1 = series.quantile(0.25)
Q3 = series.quantile(0.75)
IQR = Q3 - Q1
mask = (series >= Q1 - 1.5 * IQR) & (series <= Q3 + 1.5 * IQR)
self.current_df = self.current_df[mask | series.isna()]
elif method == "zscore":
mean = series.mean()
std = series.std() + 1e-9
mask = ((series - mean).abs() <= 3 * std)
self.current_df = self.current_df[mask | series.isna()]
self.current_df = self.current_df.reset_index(drop=True)
messages.append(c)
removed = before - len(self.current_df)
return f"Removed {removed} outlier rows from: {', '.join(messages)}"
def _op_rename_columns(self, params: Dict) -> str:
mapping = params.get("mapping", {})
if not mapping:
# Auto snake_case
new_names = {
col: col.lower().replace(" ", "_")
for col in self.current_df.columns
}
self.current_df = self.current_df.rename(columns=new_names)
return f"Auto renamed columns to snake_case: {list(new_names.values())}"
self.current_df = self.current_df.rename(columns=mapping)
return f"Renamed columns: {mapping}"
def _op_validate_schema(self, params: Dict) -> str:
gold_cols = list(self.gold_df.columns)
curr_cols = list(self.current_df.columns)
missing = [c for c in gold_cols if c not in curr_cols]
extra = [c for c in curr_cols if c not in gold_cols]
if not missing and not extra:
return "Schema valid! All columns match gold standard."
msg = []
if missing:
msg.append(f"Missing columns: {missing}")
if extra:
msg.append(f"Extra columns: {extra}")
return "Schema issues: " + " | ".join(msg)
# ─────────────────────────────────────────
# OPENENV API
# ─────────────────────────────────────────
def reset(self) -> StepResult:
self._load_datasets()
self.step_count = 0
self.done = False
self.reward_history = []
self.actions_taken = [] # Reset action history
obs = self._get_observation("Environment reset. Start cleaning!")
reward = Reward(total=0.0001)
return StepResult(
observation=obs,
reward=reward,
done=False,
info={"task_id": self.task_id, "max_steps": self.max_steps}
)
def step(self, action: Action) -> StepResult:
if self.done:
obs = self._get_observation("Episode already done. Call reset().")
return StepResult(
observation=obs,
reward=Reward(total=0.0001),
done=True,
info={"warning": "Episode already done"}
)
self.step_count += 1
op = action.operation
params = action.parameters
message = ""
# Map fill_missing_* operations to fill_missing with strategy parameter
if op.startswith("fill_missing_"):
strategy = op.replace("fill_missing_", "")
params["strategy"] = strategy
op = "fill_missing"
# ── Route operation ──────────────────────────────────────────
try:
if op == "remove_duplicates":
message = self._op_remove_duplicates(params)
elif op == "fill_missing":
message = self._op_fill_missing(params)
elif op == "fix_dtype":
message = self._op_fix_dtype(params)
elif op == "remove_outliers":
message = self._op_remove_outliers(params)
elif op == "rename_columns":
message = self._op_rename_columns(params)
elif op == "validate_schema":
message = self._op_validate_schema(params)
elif op == "finish":
message = "Agent called finish."
self.done = True
else:
message = f"Unknown operation: {op}. No changes made."
except Exception as e:
message = f"Operation failed: {str(e)}"
# ── Track action for sequence penalties ───────────────────────────
self.actions_taken.append(action.operation)
# ── Check max steps ──────────────────────────────────────────
if self.step_count >= self.max_steps:
self.done = True
message += " | Max steps reached."
reward = self._compute_reward()
self.reward_history.append(reward.total)
obs = self._get_observation(message)
return StepResult(
observation=obs,
reward=reward,
done=self.done,
info={
"step": self.step_count,
"operation": action.operation,
"reward_history": self.reward_history
}
)
def state(self) -> Dict[str, Any]:
if self.current_df is None:
return {"status": "not initialized β€” call reset() first"}
return {
"task_id": self.task_id,
"step": self.step_count,
"done": self.done,
"shape": list(self.current_df.shape),
"columns": list(self.current_df.columns),
"missing_values": self.current_df.isnull().sum().to_dict(),
"duplicate_count": int(self.current_df.duplicated().sum()),
"reward_history": self.reward_history,
"dtypes": {c: str(t) for c, t in self.current_df.dtypes.items()}
}