Datasets:
File size: 16,423 Bytes
06c7e2f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 | """Base task class and configuration. When implementing a new task, please put it under tasks/implementations and inherit from BaseTask."""
import json
import csv
import pandas as pd
from abc import ABC, abstractmethod
from dataclasses import dataclass, field
from typing import Dict, List, Any, Optional, Union, Callable
from pathlib import Path
@dataclass
class TaskConfig:
"""Configuration for a task."""
name: str # the task name
description: Optional[str] = None # optional description
data_path: Optional[str] = None # path to the main data file (CSV, JSON, etc.) for non-memory tasks
data_format: str = "csv" # 'csv', 'json', 'jsonl', 'memory'
input_column: str = "input" # name of the input column in the data
output_column: str = "output" # name of the output column in the data
demonstrations_path: Optional[str] = None # path to demonstrations file (JSON, CSV, TXT)
num_demonstrations: int = 5 # number of demonstrations to include in prompt. Will be max(num available, num_demonstrations)
prompt_template: Optional[str] = None # optional prompt template with placeholders
evaluation_metrics: List[str] = field(default_factory=lambda: ["accuracy"])
metadata: Dict[str, Any] = field(default_factory=dict)
# For in-memory data (when migrating from hardcoded tasks)
in_memory_data: Optional[List[Dict[str, Any]]] = None # option to just pass in-memory data (e.g. just a list)
in_memory_demonstrations: Optional[Dict[str, List[str]]] = None
# Optional override for the number of new tokens to generate during eval.
# When set on the task, evaluate_model() will use this value instead of its
# own max_new_tokens argument (which comes from the CLI default of 100).
max_new_tokens: Optional[int] = None
class BaseTask(ABC):
"""Base class for all tasks"""
def __init__(self, config: TaskConfig):
self.config = config
self.data = None
self.demonstrations = None
# track which example indices have been used for ICL sampling
self._icl_used_indices = set()
self._load_data()
self._load_demonstrations()
def _load_data(self):
"""Load the main task data from the specified path and format or use in-memory data."""
# Check if we have in-memory data first
if self.config.in_memory_data is not None:
self.data = pd.DataFrame(self.config.in_memory_data)
return
# If no data_path provided, let subclass handle data loading
if not self.config.data_path:
# Subclasses should override this method to provide their own data
# If they don't override it properly, self.data will remain None
return
# Otherwise load from file
data_path = Path(self.config.data_path)
if not data_path.exists():
raise FileNotFoundError(f"Data file not found: {data_path}")
if self.config.data_format.lower() == 'csv':
self.data = pd.read_csv(data_path)
elif self.config.data_format.lower() == 'json':
with open(data_path, 'r', encoding='utf-8') as f:
data = json.load(f)
self.data = pd.DataFrame(data)
elif self.config.data_format.lower() == 'jsonl':
data = []
with open(data_path, 'r', encoding='utf-8') as f:
for line in f:
data.append(json.loads(line.strip()))
self.data = pd.DataFrame(data)
else:
raise ValueError(f"Unsupported data format: {self.config.data_format}")
def _load_demonstrations(self):
"""Load demonstration examples from file, in-memory data, or use defaults."""
# Check for in-memory demonstrations first
if self.config.in_memory_demonstrations is not None:
self.demonstrations = self.config.in_memory_demonstrations
return
# Otherwise try to load from file
if self.config.demonstrations_path:
demo_path = Path(self.config.demonstrations_path)
if demo_path.exists():
if demo_path.suffix == '.json':
with open(demo_path, 'r', encoding='utf-8') as f:
self.demonstrations = json.load(f)
elif demo_path.suffix == '.csv':
self.demonstrations = pd.read_csv(demo_path)
else:
# Try to load as text file with examples
with open(demo_path, 'r', encoding='utf-8') as f:
self.demonstrations = f.read().strip()
@property
def supports_icl(self) -> bool:
"""Check if this task supports ICL format"""
return hasattr(self, 'get_icl_examples') or self.config.in_memory_data is not None
def get_icl_examples(
self,
num_examples: int = 10,
shuffle: bool = True,
seed: Optional[int] = None,
fresh: bool = True,
) -> List[Dict[str, str]]:
"""Return ICL-formatted examples (dicts with 'input' and 'output').
Args:
num_examples: number of examples to return.
shuffle: whether to shuffle available examples before selection.
seed: optional random seed for reproducibility.
fresh: if True, prefer examples that haven't been returned before
(tracks indices in ``self._icl_used_indices``).
"""
rows = self.get_split("test")
if not rows:
return []
n = len(rows)
indices = list(range(n))
if shuffle:
import random
if seed is not None:
random.seed(seed)
random.shuffle(indices)
if fresh:
# choose indices not seen before; if insufficient, wrap around deterministically
available = [i for i in indices if i not in self._icl_used_indices]
if len(available) < num_examples:
# include previously used indices to fill up
available = available + [i for i in indices if i not in available]
selected = available[:min(num_examples, n)]
self._icl_used_indices.update(selected)
else:
selected = indices[:min(num_examples, n)]
examples: List[Dict[str, str]] = []
for i in selected:
item = rows[i]
examples.append({
"input": item.get(self.config.input_column, ""),
"output": str(item.get(self.config.output_column, "")),
})
return examples
def reset_icl_tracking(self) -> None:
"""Reset internal tracking of used ICL examples so sampling starts fresh."""
self._icl_used_indices.clear()
def get_split(self, split: str = "test") -> List[Dict[str, Any]]:
"""Get data split as a list of dictionaries."""
if split not in ["train", "val", "test", "all"]:
raise ValueError(f"Invalid split: {split}. Must be one of ['train', 'val', 'test', 'all']")
# For now, return all data as test split
# Subclasses can override this for proper train/val/test splits
if split == "all" or split == "test":
# Handle both DataFrame and list data
if hasattr(self.data, 'to_dict'):
return self.data.to_dict('records')
elif isinstance(self.data, list):
return self.data
else:
return list(self.data)
else:
# Return empty for train/val if not implemented
return []
def build_prompt(self, instance: Dict[str, Any], num_shots: int = 5) -> str:
"""Build a prompt for a given instance.
Args:
instance: The instance to build a prompt for
num_shots: Number of in-context learning examples to include (default: 5)
Returns:
The formatted prompt string
"""
prompt = ""
# Add demonstrations if available
if self.demonstrations is not None:
prompt += self._format_demonstrations()
prompt += "\n\n"
# If task supports ICL generation, use that
elif hasattr(self, 'get_icl_examples') and num_shots > 0:
try:
icl_examples = self.get_icl_examples(num_examples=num_shots)
if icl_examples:
prompt += self._format_icl_examples(icl_examples)
prompt += "\n\n"
except Exception as e:
# If ICL generation fails, continue without examples
pass
# Add the instance input
if self.config.prompt_template:
prompt += self.config.prompt_template.format(**instance)
else:
prompt += f"Input: {instance[self.config.input_column]}\nOutput:"
return prompt
def _format_demonstrations(self) -> str:
"""Format demonstration examples into a string."""
if isinstance(self.demonstrations, str):
return self.demonstrations
elif isinstance(self.demonstrations, pd.DataFrame):
demo_text = ""
for _, row in self.demonstrations.head(self.config.num_demonstrations).iterrows():
demo_text += f"Input: {row[self.config.input_column]}\nOutput: {row[self.config.output_column]}\n\n"
return demo_text.strip()
elif isinstance(self.demonstrations, list):
demo_text = ""
for demo in self.demonstrations[:self.config.num_demonstrations]:
demo_text += f"Input: {demo[self.config.input_column]}\nOutput: {demo[self.config.output_column]}\n\n"
return demo_text.strip()
elif isinstance(self.demonstrations, dict):
# Handle category-based demonstrations (like in the notebook)
return "" # Will be handled by subclasses that need category-specific demos
else:
return ""
def _format_icl_examples(self, examples: List[Dict[str, Any]]) -> str:
"""Format ICL examples into a string.
Args:
examples: List of example dictionaries with input/output keys
Returns:
Formatted string of examples
"""
if not examples:
return ""
demo_text = "\n"
for ex in examples:
# Try different column names
input_val = ex.get(self.config.input_column) or ex.get('input') or ex.get('question')
output_val = ex.get(self.config.output_column) or ex.get('output') or ex.get('answer')
if input_val is not None and output_val is not None:
demo_text += f"Input: {input_val}\nOutput: {output_val}\n\n"
return demo_text.strip()
@abstractmethod
def evaluate(self, predictions: List[str], split: str = "test", **kwargs) -> Dict[str, float]:
"""Evaluate predictions against ground truth."""
pass
def preprocess_prediction(self, prediction: str) -> str:
"""Preprocess a model prediction before evaluation."""
# Default implementation: strip whitespace and take first line
return prediction.strip().split('\n')[0] if prediction else ""
def get_ground_truth(self, split: str = "test") -> List[str]:
"""Get ground truth answers for a split."""
data = self.get_split(split)
return [item[self.config.output_column] for item in data]
class SimpleTask(BaseTask):
"""A simple task implementation with basic exact match evaluation."""
def evaluate(self, predictions: List[str], split: str = "test", **kwargs) -> Dict[str, float]:
"""Evaluate predictions with exact match accuracy."""
ground_truth = self.get_ground_truth(split)
if len(predictions) != len(ground_truth):
raise ValueError(f"Prediction count ({len(predictions)}) doesn't match ground truth count ({len(ground_truth)})")
# Preprocess predictions
processed_predictions = [self.preprocess_prediction(pred) for pred in predictions]
# Calculate exact match accuracy
correct = sum(1 for pred, gt in zip(processed_predictions, ground_truth)
if pred.lower().strip() == gt.lower().strip())
accuracy = correct / len(ground_truth)
results = {
"accuracy": accuracy,
"correct": correct,
"total": len(ground_truth)
}
# Add per-category accuracy if available
data = self.get_split(split)
if any("category" in item for item in data):
category_results = {}
for pred, gt, item in zip(processed_predictions, ground_truth, data):
category = item.get("category", "unknown")
if category not in category_results:
category_results[category] = {"correct": 0, "total": 0}
category_results[category]["total"] += 1
if pred.lower().strip() == gt.lower().strip():
category_results[category]["correct"] += 1
# Calculate per-category accuracy
for category, stats in category_results.items():
results[f"accuracy_{category}"] = stats["correct"] / stats["total"]
return results
# Factory function for creating tasks from config files
def create_task_from_config(config_path: str, task_class: type = SimpleTask) -> BaseTask:
"""Create a task instance from a configuration file."""
config_path = Path(config_path)
if config_path.suffix == '.json':
with open(config_path, 'r', encoding='utf-8') as f:
config_dict = json.load(f)
else:
raise ValueError(f"Unsupported config format: {config_path.suffix}")
config = TaskConfig(**config_dict)
return task_class(config)
# Factory function for creating tasks from hardcoded data (for migration)
def create_task_from_data(
name: str,
data: List[Dict[str, Any]],
demonstrations: Optional[Dict[str, List[str]]] = None,
description: str = "Task created from hardcoded data",
task_class: type = SimpleTask,
**kwargs
) -> BaseTask:
"""Create a task instance from in-memory data."""
config = TaskConfig(
name=name,
description=description,
data_format="memory",
in_memory_data=data,
in_memory_demonstrations=demonstrations,
**kwargs
)
return task_class(config)
# Utility function to convert notebook-style data to standardized format
def convert_notebook_data_to_standard(
csv_path: str,
category_demonstrations: Dict[str, List[str]],
output_csv_path: Optional[str] = None,
output_demo_json_path: Optional[str] = None
) -> tuple[str, str]:
"""
Convert notebook-style hardcoded data to standardized CSV and JSON files.
Args:
csv_path: Path to the existing CSV data file
category_demonstrations: Dictionary mapping categories to demonstration examples
output_csv_path: Path for output CSV (defaults to input path)
output_demo_json_path: Path for demonstration JSON file
Returns:
Tuple of (csv_path, demo_json_path)
"""
import pandas as pd
import json
from pathlib import Path
# Read the CSV data
data = pd.read_csv(csv_path)
# Set default output paths
if output_csv_path is None:
output_csv_path = csv_path
if output_demo_json_path is None:
base_path = Path(csv_path).parent
output_demo_json_path = base_path / f"{Path(csv_path).stem}_demonstrations.json"
# Save demonstrations to JSON
with open(output_demo_json_path, 'w', encoding='utf-8') as f:
json.dump(category_demonstrations, f, indent=2, ensure_ascii=False)
# Ensure CSV has the right format (it likely already does)
if output_csv_path != csv_path:
data.to_csv(output_csv_path, index=False)
print(f"Converted data saved to:")
print(f" CSV: {output_csv_path}")
print(f" Demonstrations: {output_demo_json_path}")
return str(output_csv_path), str(output_demo_json_path)
|