mirror of
https://github.com/open-thought/reasoning-gym.git
synced 2026-04-19 12:58:07 +00:00
Refactor LetterJumble
This commit is contained in:
EduardDurech
parent
b8ce5a8a5d
commit
18b6e71fa9
6 changed files with 550 additions and 190 deletions
105
examples/exercises/algorithmic/letter_jumble_examples.py
Normal file
105
examples/exercises/algorithmic/letter_jumble_examples.py
Normal file
|
|
@ -0,0 +1,105 @@
|
|||
"""Examples of generated problems from the LetterJumble exercise.
|
||||
|
||||
This file demonstrates different types of letter jumble problems that can be generated
|
||||
at various difficulty levels.
|
||||
"""
|
||||
|
||||
import random
|
||||
from reasoning_gym.curricula.algorithmic.letter_jumble_curriculum import LetterJumbleCurriculum
|
||||
from reasoning_gym.exercises.algorithmic.letter_jumble import LetterJumbleExercise
|
||||
|
||||
def main():
|
||||
# Initialize with fixed seed for reproducibility
|
||||
curriculum = LetterJumbleCurriculum()
|
||||
exercise = LetterJumbleExercise()
|
||||
curriculum.rng = random.Random(42)
|
||||
|
||||
print("\n========================================\n")
|
||||
|
||||
# Level 0: Basic word scrambling
|
||||
curriculum.set_attr_level("word_length", 0) # Short words (up to 5 chars)
|
||||
curriculum.set_attr_level("num_words", 0) # Few words (up to 3)
|
||||
curriculum.set_attr_level("corruption_level", 0) # Light scrambling (0.3)
|
||||
curriculum.set_attr_level("consecutive_words", 0) # Consecutive words
|
||||
curriculum.set_attr_level("preserve_length", 0) # Preserve first 4 chars
|
||||
problem = exercise.generate(curriculum)
|
||||
print("Level 0 (Basic Word Scrambling):")
|
||||
print(problem)
|
||||
|
||||
print("\n========================================\n")
|
||||
|
||||
# Level 1: Medium difficulty
|
||||
curriculum.set_attr_level("word_length", 1) # Medium words (up to 8 chars)
|
||||
curriculum.set_attr_level("num_words", 1) # More words (up to 5)
|
||||
curriculum.set_attr_level("corruption_level", 1) # Medium scrambling (0.6)
|
||||
curriculum.set_attr_level("consecutive_words", 0) # Consecutive words
|
||||
curriculum.set_attr_level("preserve_length", 0) # Preserve first 4 chars
|
||||
problem = exercise.generate(curriculum)
|
||||
print("Level 1 (Medium Difficulty):")
|
||||
print(problem)
|
||||
|
||||
print("\n========================================\n")
|
||||
|
||||
# Level 2: Advanced scrambling
|
||||
curriculum.set_attr_level("word_length", 2) # Long words (up to 64 chars)
|
||||
curriculum.set_attr_level("num_words", 2) # Many words (up to 20)
|
||||
curriculum.set_attr_level("corruption_level", 2) # Heavy scrambling (0.9)
|
||||
curriculum.set_attr_level("consecutive_words", 1) # Non-consecutive words
|
||||
curriculum.set_attr_level("preserve_length", 1) # Preserve first 2 chars
|
||||
problem = exercise.generate(curriculum)
|
||||
print("Level 2 (Advanced Scrambling):")
|
||||
print(problem)
|
||||
|
||||
print("\n========================================\n")
|
||||
|
||||
# Random Examples with Different Seeds
|
||||
print("Random Examples (Different Seeds):")
|
||||
for seed in range(10, 15):
|
||||
curriculum.rng = random.Random(seed)
|
||||
# Randomly set curriculum levels
|
||||
curriculum.set_attr_level("word_length", random.randint(0, 2))
|
||||
curriculum.set_attr_level("num_words", random.randint(0, 2))
|
||||
curriculum.set_attr_level("corruption_level", random.randint(0, 2))
|
||||
curriculum.set_attr_level("consecutive_words", random.randint(0, 1))
|
||||
curriculum.set_attr_level("preserve_length", random.randint(0, 1))
|
||||
problem = exercise.generate(curriculum)
|
||||
print(f"\nRandom Example (Seed {seed}):")
|
||||
print(problem)
|
||||
|
||||
print("\n========================================\n")
|
||||
|
||||
# Special Cases
|
||||
print("Special Cases:")
|
||||
|
||||
# Case 1: Maximum length single word with minimal preservation
|
||||
curriculum.set_attr_level("word_length", 2) # Long words
|
||||
curriculum.set_attr_level("num_words", 0) # Single word
|
||||
curriculum.set_attr_level("corruption_level", 2) # Heavy scrambling
|
||||
curriculum.set_attr_level("consecutive_words", 0) # Consecutive (doesn't matter for single word)
|
||||
curriculum.set_attr_level("preserve_length", 1) # Preserve first 2 chars
|
||||
problem = exercise.generate(curriculum)
|
||||
print("\nLong Single Word (Minimal Preservation):")
|
||||
print(problem)
|
||||
|
||||
# Case 2: Many short words with maximum preservation
|
||||
curriculum.set_attr_level("word_length", 0) # Short words
|
||||
curriculum.set_attr_level("num_words", 2) # Many words
|
||||
curriculum.set_attr_level("corruption_level", 1) # Medium scrambling
|
||||
curriculum.set_attr_level("consecutive_words", 1) # Non-consecutive
|
||||
curriculum.set_attr_level("preserve_length", 0) # Preserve first 4 chars
|
||||
problem = exercise.generate(curriculum)
|
||||
print("\nMany Short Words (Maximum Preservation):")
|
||||
print(problem)
|
||||
|
||||
# Case 3: Medium words with balanced preservation
|
||||
curriculum.set_attr_level("word_length", 1) # Medium words
|
||||
curriculum.set_attr_level("num_words", 1) # Medium number of words
|
||||
curriculum.set_attr_level("corruption_level", 0) # Light scrambling
|
||||
curriculum.set_attr_level("consecutive_words", 0) # Consecutive
|
||||
curriculum.set_attr_level("preserve_length", 1) # Preserve first 2 chars
|
||||
problem = exercise.generate(curriculum)
|
||||
print("\nMedium Words (Balanced Preservation):")
|
||||
print(problem)
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
|
|
@ -1,103 +1,66 @@
|
|||
"""Word letter jumbling task generator"""
|
||||
"""Exercise definition for letter jumble exercises."""
|
||||
|
||||
import re
|
||||
from dataclasses import dataclass
|
||||
from random import Random
|
||||
from typing import List, Optional
|
||||
from typing import Dict, Any
|
||||
from reasoning_gym.core.template import Template
|
||||
|
||||
from reasoning_gym.data import read_data_file
|
||||
class LetterJumbleExercise:
|
||||
"""Exercise generator for word jumbling tasks."""
|
||||
|
||||
from ..factory import ProceduralDataset, register_dataset
|
||||
def __init__(self):
|
||||
self.curriculum = None
|
||||
|
||||
def generate(self, curriculum: Any) -> Dict[str, Any]:
|
||||
"""
|
||||
Generate a word jumbling problem using the curriculum.
|
||||
|
||||
@dataclass
|
||||
class LetterJumbleConfig:
|
||||
"""Configuration for letter jumbling task generation"""
|
||||
Returns:
|
||||
Dict containing:
|
||||
- question: str (e.g. "Unscramble these words: OLHEL DLWOR")
|
||||
- answer: str (the original words)
|
||||
- metadata: dict with details (scrambled_words, original_words, etc.)
|
||||
"""
|
||||
self.curriculum = curriculum
|
||||
template = curriculum.get_template(curriculum.rng)
|
||||
return template.eval(self, curriculum.rng)
|
||||
|
||||
min_word_len: int = 1 # Minimum word length
|
||||
max_word_len: int = 64 # Maximum word length
|
||||
min_words: int = 3 # Minimum words per task
|
||||
max_words: int = 20 # Maximum words per task
|
||||
min_corruption_level: float = 0.1 # Minimum fraction of characters to swap
|
||||
max_corruption_level: float = 0.9 # Maximum fraction of characters to swap
|
||||
consecutive_words: bool = True # Whether to select consecutive words from text
|
||||
seed: Optional[int] = None
|
||||
size: int = 500 # Virtual dataset size
|
||||
def _parse_expression(self, metadata: Dict[str, Any]) -> Dict[str, Any]:
|
||||
"""Parse the expression from the metadata.
|
||||
|
||||
def validate(self) -> None:
|
||||
"""Validate configuration parameters"""
|
||||
assert self.min_word_len > 0, "min_word_len must be positive"
|
||||
assert self.max_word_len >= self.min_word_len, "max_word_len must be >= min_word_len"
|
||||
assert self.min_words > 0, "min_words must be positive"
|
||||
assert self.max_words >= self.min_words, "max_words must be >= min_words"
|
||||
assert 0 <= self.min_corruption_level <= 1, "min_corruption_level must be in [0,1]"
|
||||
assert 0 <= self.max_corruption_level <= 1, "max_corruption_level must be in [0,1]"
|
||||
assert (
|
||||
self.max_corruption_level >= self.min_corruption_level
|
||||
), "max_corruption_level must be >= min_corruption_level"
|
||||
|
||||
|
||||
class LetterJumbleDataset(ProceduralDataset):
|
||||
"""Generates word letter jumbling tasks"""
|
||||
|
||||
def __init__(self, config: LetterJumbleConfig):
|
||||
super().__init__(config=config, seed=config.seed, size=config.size)
|
||||
|
||||
# Load and preprocess text
|
||||
text = read_data_file("in_the_year_2889.txt")
|
||||
# Extract words and filter by length
|
||||
self.words = [
|
||||
word
|
||||
for word in re.findall(r"\b\w+\b", text)
|
||||
if self.config.min_word_len <= len(word) <= self.config.max_word_len and word.isalpha()
|
||||
]
|
||||
|
||||
def _scramble_word(self, word: str, corruption_level: float, rng: Random) -> str:
|
||||
"""Scramble a word by swapping random pairs of characters"""
|
||||
if len(word) < 2: # Can't scramble 1-character words
|
||||
return word
|
||||
|
||||
word = list(word)
|
||||
num_swaps = max(1, int(len(word) * corruption_level)) # Ensure at least one swap
|
||||
|
||||
for _ in range(num_swaps):
|
||||
# Pick two different random positions
|
||||
pos1, pos2 = rng.sample(range(len(word)), 2)
|
||||
# Swap characters
|
||||
word[pos1], word[pos2] = word[pos2], word[pos1]
|
||||
|
||||
return "".join(word)
|
||||
|
||||
def __getitem__(self, idx: int) -> dict:
|
||||
"""Generate a single word jumbling task"""
|
||||
rng = Random(self.seed + idx)
|
||||
|
||||
# Select number of words and corruption level
|
||||
num_words = rng.randint(self.config.min_words, self.config.max_words)
|
||||
corruption_level = rng.uniform(self.config.min_corruption_level, self.config.max_corruption_level)
|
||||
|
||||
# Select words based on configuration
|
||||
if self.config.consecutive_words:
|
||||
# Select consecutive words from a random starting position
|
||||
start_idx = rng.randint(0, len(self.words) - num_words)
|
||||
selected_words = self.words[start_idx : start_idx + num_words]
|
||||
else:
|
||||
# Select random words
|
||||
selected_words = rng.sample(self.words, num_words)
|
||||
|
||||
# Scramble each word
|
||||
scrambled_words = [self._scramble_word(word, corruption_level, rng) for word in selected_words]
|
||||
|
||||
return {
|
||||
"question": f"Unscramble these words: {' '.join(scrambled_words)}",
|
||||
"answer": " ".join(selected_words),
|
||||
"metadata": {
|
||||
"num_words": num_words,
|
||||
"corruption_level": corruption_level,
|
||||
"scrambled_words": scrambled_words,
|
||||
"original_words": selected_words,
|
||||
},
|
||||
The metadata structure from the template system:
|
||||
{
|
||||
"scrambled": {
|
||||
"scrambled_words": str, # Space-separated scrambled words
|
||||
"original_words": List[str] # List of original words
|
||||
}
|
||||
}
|
||||
|
||||
Args:
|
||||
metadata: The metadata containing the expression information.
|
||||
|
||||
register_dataset("letter_jumble", LetterJumbleDataset, LetterJumbleConfig)
|
||||
Returns:
|
||||
A dictionary containing:
|
||||
- scrambled_words: List[str] of scrambled words
|
||||
- original_words: List[str] of original words
|
||||
"""
|
||||
# Extract the scrambled and original words from metadata
|
||||
template_data = metadata["scrambled"]
|
||||
scrambled_words = template_data["scrambled_words"].split()
|
||||
original_words = template_data["original_words"]
|
||||
|
||||
return {
|
||||
"scrambled_words": scrambled_words,
|
||||
"original_words": original_words
|
||||
}
|
||||
|
||||
def _evaluate_expression(self, parsed_data: Dict[str, Any]) -> str:
|
||||
"""Evaluate the expression using the parsed data.
|
||||
|
||||
Args:
|
||||
parsed_data: Dictionary containing:
|
||||
- scrambled_words: List[str] of scrambled words
|
||||
- original_words: List[str] of original words
|
||||
|
||||
Returns:
|
||||
The answer string (space-separated original words).
|
||||
"""
|
||||
return " ".join(parsed_data["original_words"])
|
||||
|
|
|
|||
|
|
@ -1,8 +1,10 @@
|
|||
from .base_conversion_curriculum import BaseConversionCurriculum
|
||||
from .caesar_cipher_curriculum import CaesarCipherCurriculum
|
||||
from .letter_counting_curriculum import LetterCountingCurriculum
|
||||
from .letter_jumble_curriculum import LetterJumbleCurriculum
|
||||
__all__ = [
|
||||
"BaseConversionCurriculum",
|
||||
"CaesarCipherCurriculum",
|
||||
"LetterCountingCurriculum"
|
||||
"LetterCountingCurriculum",
|
||||
"LetterJumbleCurriculum"
|
||||
]
|
||||
|
|
|
|||
122
reasoning_gym/curricula/algorithmic/letter_jumble_curriculum.py
Normal file
122
reasoning_gym/curricula/algorithmic/letter_jumble_curriculum.py
Normal file
|
|
@ -0,0 +1,122 @@
|
|||
"""
|
||||
Curriculum definition for letter jumble exercises.
|
||||
"""
|
||||
|
||||
from typing import Dict, Any
|
||||
from reasoning_gym.core.base_curriculum import BaseCurriculum
|
||||
from reasoning_gym.core.attributes import AttributeDefinition, AttributeType
|
||||
from reasoning_gym.core.template import Template
|
||||
from reasoning_gym.data import read_data_file
|
||||
|
||||
class LetterJumbleCurriculum(BaseCurriculum):
|
||||
def __init__(self):
|
||||
super().__init__("LetterJumbleCurriculum")
|
||||
import re
|
||||
self.words = [word for word in re.findall(r"\b\w+\b", read_data_file("in_the_year_2889.txt")) if word.isalpha()]
|
||||
|
||||
def _init_curriculum(self) -> None:
|
||||
"""Initialize the letter jumble curriculum configuration"""
|
||||
# Define valid attribute types
|
||||
self._valid_types = {
|
||||
AttributeType.STATIC, # For boolean flags
|
||||
AttributeType.UBOUND, # For ranges like word length, num words
|
||||
AttributeType.APPEND # For accumulating options
|
||||
}
|
||||
|
||||
# Define attributes
|
||||
self._attributes = {
|
||||
"word_length": AttributeDefinition(
|
||||
levels=[7, 12, 64], # From min_word_len/max_word_len
|
||||
default_level=0,
|
||||
description="Maximum word length",
|
||||
attr_type=AttributeType.UBOUND,
|
||||
min_value=1 # Ensure at least 2 chars for scrambling
|
||||
),
|
||||
"preserve_length": AttributeDefinition(
|
||||
levels=[4, 2],
|
||||
default_level=0,
|
||||
description="Word length to preserve",
|
||||
attr_type=AttributeType.STATIC
|
||||
),
|
||||
"num_words": AttributeDefinition(
|
||||
levels=[3, 5, 20], # From min_words/max_words
|
||||
default_level=0,
|
||||
description="Number of words to scramble",
|
||||
attr_type=AttributeType.UBOUND,
|
||||
min_value=1 # Ensure at least 1 word
|
||||
),
|
||||
"corruption_level": AttributeDefinition(
|
||||
levels=[0.1, 0.3, 0.9], # From min/max_corruption_level
|
||||
default_level=0,
|
||||
description="Fraction of characters to swap",
|
||||
attr_type=AttributeType.UBOUND,
|
||||
min_value=0.1
|
||||
),
|
||||
"consecutive_words": AttributeDefinition(
|
||||
levels=[True, False],
|
||||
default_level=0,
|
||||
description="Whether to select consecutive words",
|
||||
attr_type=AttributeType.APPEND
|
||||
)
|
||||
}
|
||||
|
||||
# Define templates with symbolic placeholders
|
||||
self._templates = [
|
||||
Template(
|
||||
template="Unscramble these words: \"{scrambled}\"",
|
||||
parts={"scrambled": "word_list"}
|
||||
),
|
||||
Template(
|
||||
template="What are the original words? \"{scrambled}\"",
|
||||
parts={"scrambled": "word_list"}
|
||||
),
|
||||
Template(
|
||||
template="Rearrange the letters to find the original words: \"{scrambled}\"",
|
||||
parts={"scrambled": "word_list"}
|
||||
)
|
||||
]
|
||||
|
||||
# Define symbolic structure
|
||||
self._symbolic = {
|
||||
# Shared variables that need to be consistent across templates
|
||||
"shared_vars": {
|
||||
# Selected original words that will be scrambled
|
||||
"selected_words": lambda refs: (
|
||||
n_words := refs["num_words"](),
|
||||
pool := self.words,
|
||||
refs["dataset_rng"].sample(pool, n_words) if not refs["consecutive_words"]() else
|
||||
(
|
||||
start := refs["dataset_rng"].randint(0, max(0, len(pool)-n_words)),
|
||||
pool[start:start + n_words]
|
||||
)[-1]
|
||||
)[-1]
|
||||
},
|
||||
# Value generators for dynamic content
|
||||
"generators": {
|
||||
# Scramble a single word based on corruption level
|
||||
"scramble_word": lambda refs: lambda lst: (
|
||||
[
|
||||
(i, j, lst.__setitem__(i, lst[j]), lst.__setitem__(j, temp)) # Debugging: keep track of indices and assignments
|
||||
for _ in range(max(0, int(len(lst) * refs["corruption_level"]())))
|
||||
for i, j in [refs["dataset_rng"].sample(range(len(lst)), 2)]
|
||||
for temp in [lst[i]] # Introduce temp variable for correct swap
|
||||
],
|
||||
"".join(lst)
|
||||
)[-1],
|
||||
# Generate scrambled version of all selected words
|
||||
"scramble_all": lambda refs: lambda: [
|
||||
refs["scramble_word"](refs)(list(word)) if len(word) > refs["preserve_length"]() else word
|
||||
for word in refs["selected_words"](refs)
|
||||
]
|
||||
},
|
||||
# Template composition
|
||||
"templates": {
|
||||
"word_list": lambda refs: {
|
||||
"template": "{scrambled_words}",
|
||||
"parts": {
|
||||
"scrambled_words": lambda refs=refs: " ".join(refs["scramble_all"](refs)()),
|
||||
"original_words": lambda refs=refs: refs["selected_words"](refs)
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -9,7 +9,7 @@ Algorithmic tasks for training reasoning capabilities:
|
|||
from .base_conversion import BaseConversionExercise
|
||||
from .caesar_cipher import CaesarCipherExercise
|
||||
from .letter_counting import LetterCountingExercise
|
||||
# from .letter_jumble import LetterJumbleExercise
|
||||
from .letter_jumble import LetterJumbleExercise
|
||||
# from .number_filtering import NumberFilteringExercise
|
||||
# from .number_sorting import NumberSortingExercise
|
||||
# from .sentence_reordering import SentenceReorderingExercise
|
||||
|
|
@ -23,7 +23,7 @@ __all__ = [
|
|||
"BaseConversionExercise",
|
||||
"CaesarCipherExercise",
|
||||
"LetterCountingExercise",
|
||||
# "LetterJumbleDataset",
|
||||
"LetterJumbleExercise",
|
||||
# "NumberFilteringDataset",
|
||||
# "NumberSortingDataset",
|
||||
# "SentenceReorderingDataset",
|
||||
|
|
|
|||
|
|
@ -1,121 +1,289 @@
|
|||
"""Tests for letter jumbling task generation"""
|
||||
"""Unit tests for the letter jumble exercise."""
|
||||
|
||||
from random import Random
|
||||
from reasoning_gym.curricula.algorithmic.letter_jumble_curriculum import LetterJumbleCurriculum
|
||||
from reasoning_gym.exercises.algorithmic.letter_jumble import LetterJumbleExercise
|
||||
import unittest
|
||||
import random
|
||||
from collections import defaultdict
|
||||
|
||||
import pytest
|
||||
class TestLetterJumbleParsing(unittest.TestCase):
|
||||
"""Test parsing of letter jumble metadata"""
|
||||
|
||||
from reasoning_gym.algorithmic.letter_jumble import LetterJumbleConfig, LetterJumbleDataset
|
||||
def setUp(self):
|
||||
self.exercise = LetterJumbleExercise()
|
||||
|
||||
def test_parse_expression_basic(self):
|
||||
"""Test parsing of basic letter jumble metadata"""
|
||||
test_metadata = {
|
||||
"scrambled": {
|
||||
"scrambled_words": "EHLLO DLWOR",
|
||||
"original_words": ["HELLO", "WORLD"]
|
||||
}
|
||||
}
|
||||
parsed = self.exercise._parse_expression(test_metadata)
|
||||
self.assertEqual(parsed["scrambled_words"], ["EHLLO", "DLWOR"])
|
||||
self.assertEqual(parsed["original_words"], ["HELLO", "WORLD"])
|
||||
|
||||
def test_letter_jumble_config_validation():
|
||||
"""Test that invalid configs raise appropriate errors"""
|
||||
with pytest.raises(AssertionError):
|
||||
config = LetterJumbleConfig(min_word_len=0)
|
||||
config.validate()
|
||||
def test_parse_with_spaces(self):
|
||||
"""Test parsing with spaces and punctuation"""
|
||||
test_metadata = {
|
||||
"scrambled": {
|
||||
"scrambled_words": "EHLLO DLWOR!",
|
||||
"original_words": ["HELLO", "WORLD!"]
|
||||
}
|
||||
}
|
||||
parsed = self.exercise._parse_expression(test_metadata)
|
||||
self.assertEqual(parsed["scrambled_words"], ["EHLLO", "DLWOR!"])
|
||||
self.assertEqual(parsed["original_words"], ["HELLO", "WORLD!"])
|
||||
|
||||
with pytest.raises(AssertionError):
|
||||
config = LetterJumbleConfig(min_words=10, max_words=5)
|
||||
config.validate()
|
||||
def test_parse_mixed_case(self):
|
||||
"""Test parsing with mixed case text"""
|
||||
test_metadata = {
|
||||
"scrambled": {
|
||||
"scrambled_words": "HeLlO WoRlD",
|
||||
"original_words": ["hElLo", "wOrLd"]
|
||||
}
|
||||
}
|
||||
parsed = self.exercise._parse_expression(test_metadata)
|
||||
self.assertEqual(parsed["scrambled_words"], ["HeLlO", "WoRlD"])
|
||||
self.assertEqual(parsed["original_words"], ["hElLo", "wOrLd"])
|
||||
|
||||
with pytest.raises(AssertionError):
|
||||
config = LetterJumbleConfig(min_corruption_level=-0.1)
|
||||
config.validate()
|
||||
class TestLetterJumbleEvaluation(unittest.TestCase):
|
||||
"""Test evaluation of letter jumble problems"""
|
||||
|
||||
with pytest.raises(AssertionError):
|
||||
config = LetterJumbleConfig(max_corruption_level=1.1)
|
||||
config.validate()
|
||||
def setUp(self):
|
||||
self.exercise = LetterJumbleExercise()
|
||||
|
||||
def test_basic_unscrambling(self):
|
||||
"""Test basic unscrambling cases"""
|
||||
test_cases = [
|
||||
(["EHLLO"], "HELLO"), # Single word
|
||||
(["EHLLO", "DLWOR"], "HELLO WORLD"), # Two words
|
||||
(["AAAA"], "AAAA"), # Same letters
|
||||
(["ZBAC"], "ABCZ"), # Sorted order
|
||||
(["HELLO"], "HELLO") # Already unscrambled
|
||||
]
|
||||
for scrambled, expected in test_cases:
|
||||
parsed = {
|
||||
"scrambled_words": scrambled,
|
||||
"original_words": expected.split()
|
||||
}
|
||||
result = self.exercise._evaluate_expression(parsed)
|
||||
self.assertEqual(result, expected)
|
||||
|
||||
def test_letter_jumble_deterministic():
|
||||
"""Test that dataset generates same items with same seed"""
|
||||
config = LetterJumbleConfig(seed=42, size=10)
|
||||
dataset1 = LetterJumbleDataset(config)
|
||||
dataset2 = LetterJumbleDataset(config)
|
||||
def test_mixed_case_unscrambling(self):
|
||||
"""Test unscrambling with mixed case"""
|
||||
test_cases = [
|
||||
(["HeLlO"], "hElLo"), # Mixed case, single word
|
||||
(["WoRlD", "HeLlO"], "wOrLd hElLo"), # Mixed case, multiple words
|
||||
(["AbCdE"], "aBcDe") # Mixed case, alternating
|
||||
]
|
||||
for scrambled, expected in test_cases:
|
||||
parsed = {
|
||||
"scrambled_words": scrambled,
|
||||
"original_words": expected.split()
|
||||
}
|
||||
result = self.exercise._evaluate_expression(parsed)
|
||||
self.assertEqual(result, expected)
|
||||
|
||||
for i in range(len(dataset1)):
|
||||
assert dataset1[i] == dataset2[i]
|
||||
def test_with_spaces_and_punctuation(self):
|
||||
"""Test unscrambling with spaces and punctuation"""
|
||||
test_cases = [
|
||||
(["EHLLO!", "DLWOR?"], "HELLO! WORLD?"),
|
||||
(["EHLLO.", "DLWOR."], "HELLO. WORLD."),
|
||||
(["EHLLO,", "DLWOR,"], "HELLO, WORLD,")
|
||||
]
|
||||
for scrambled, expected in test_cases:
|
||||
parsed = {
|
||||
"scrambled_words": scrambled,
|
||||
"original_words": expected.split()
|
||||
}
|
||||
result = self.exercise._evaluate_expression(parsed)
|
||||
self.assertEqual(result, expected)
|
||||
|
||||
class TestLetterJumbleGeneration(unittest.TestCase):
|
||||
"""Test problem generation"""
|
||||
|
||||
def test_letter_jumble_scrambling():
|
||||
"""Test the word scrambling logic"""
|
||||
config = LetterJumbleConfig(
|
||||
min_word_len=4,
|
||||
max_word_len=8,
|
||||
min_words=1,
|
||||
max_words=1,
|
||||
min_corruption_level=0.5,
|
||||
max_corruption_level=0.5,
|
||||
size=1,
|
||||
seed=42,
|
||||
)
|
||||
dataset = LetterJumbleDataset(config)
|
||||
def setUp(self):
|
||||
self.curriculum = LetterJumbleCurriculum()
|
||||
self.exercise = LetterJumbleExercise()
|
||||
self.rng = random.Random(42)
|
||||
self.curriculum.rng = self.rng
|
||||
|
||||
# Test with known word
|
||||
word = "testing"
|
||||
rng = Random(42)
|
||||
scrambled = dataset._scramble_word(word, 0.5, rng)
|
||||
def test_problem_structure(self):
|
||||
"""Test that generated problems have the correct structure"""
|
||||
problem = self.exercise.generate(self.curriculum)
|
||||
|
||||
# Verify scrambled word:
|
||||
# - Has same length as original
|
||||
assert len(scrambled) == len(word)
|
||||
# - Contains same characters
|
||||
assert sorted(scrambled) == sorted(word)
|
||||
# - Is different from original (with high probability given 0.5 corruption)
|
||||
assert scrambled != word
|
||||
# Check basic structure
|
||||
self.assertIn("question", problem)
|
||||
self.assertIn("answer", problem)
|
||||
self.assertIn("metadata", problem)
|
||||
|
||||
# Check metadata structure
|
||||
metadata = problem["metadata"]
|
||||
self.assertEqual(metadata["type"], "direct")
|
||||
self.assertIn("executed_parts", metadata)
|
||||
executed_parts = metadata["executed_parts"]
|
||||
self.assertIn("scrambled_words", executed_parts)
|
||||
self.assertIn("original_words", executed_parts)
|
||||
|
||||
def test_letter_jumble_dataset_items():
|
||||
"""Test basic properties of generated items"""
|
||||
config = LetterJumbleConfig(
|
||||
min_word_len=4,
|
||||
max_word_len=8,
|
||||
min_words=3,
|
||||
max_words=5,
|
||||
min_corruption_level=0.1,
|
||||
max_corruption_level=0.3,
|
||||
size=50,
|
||||
seed=42,
|
||||
)
|
||||
dataset = LetterJumbleDataset(config)
|
||||
def test_word_length_ranges(self):
|
||||
"""Test that word lengths are within expected ranges"""
|
||||
# Test all word length levels
|
||||
level_max_lengths = {0: 5, 1: 8, 2: 64}
|
||||
|
||||
for i in range(len(dataset)):
|
||||
item = dataset[i]
|
||||
for level, max_length in level_max_lengths.items():
|
||||
self.curriculum.set_attr_level("word_length", level)
|
||||
problem = self.exercise.generate(self.curriculum)
|
||||
words = problem["metadata"]["executed_parts"]["original_words"]
|
||||
for word in words:
|
||||
self.assertLessEqual(len(word), max_length)
|
||||
self.assertGreaterEqual(len(word), 2) # Min length is 2
|
||||
|
||||
# Check item structure
|
||||
assert isinstance(item, dict)
|
||||
assert "question" in item
|
||||
assert "answer" in item
|
||||
assert "metadata" in item
|
||||
def test_word_count_ranges(self):
|
||||
"""Test that word counts are within expected ranges"""
|
||||
# Test all word count levels
|
||||
level_word_counts = {0: 3, 1: 5, 2: 20}
|
||||
|
||||
# Check metadata
|
||||
metadata = item["metadata"]
|
||||
assert "num_words" in metadata
|
||||
assert "corruption_level" in metadata
|
||||
assert "scrambled_words" in metadata
|
||||
assert "original_words" in metadata
|
||||
for level, max_words in level_word_counts.items():
|
||||
self.curriculum.set_attr_level("num_words", level)
|
||||
problem = self.exercise.generate(self.curriculum)
|
||||
words = problem["metadata"]["executed_parts"]["original_words"]
|
||||
self.assertLessEqual(len(words), max_words)
|
||||
self.assertGreaterEqual(len(words), 1) # Min words is 1
|
||||
|
||||
# Verify word counts
|
||||
num_words = metadata["num_words"]
|
||||
assert config.min_words <= num_words <= config.max_words
|
||||
assert len(metadata["scrambled_words"]) == num_words
|
||||
assert len(metadata["original_words"]) == num_words
|
||||
class TestLetterJumbleComprehensive(unittest.TestCase):
|
||||
"""Comprehensive tests for letter jumble"""
|
||||
|
||||
# Verify corruption level
|
||||
assert config.min_corruption_level <= metadata["corruption_level"] <= config.max_corruption_level
|
||||
def setUp(self):
|
||||
self.curriculum = LetterJumbleCurriculum()
|
||||
self.exercise = LetterJumbleExercise()
|
||||
self.rng = random.Random(42)
|
||||
self.curriculum.rng = self.rng
|
||||
|
||||
# Verify word properties
|
||||
for word in metadata["original_words"]:
|
||||
assert config.min_word_len <= len(word) <= config.max_word_len
|
||||
assert word.isalpha()
|
||||
def test_corruption_levels(self):
|
||||
"""Test different corruption levels"""
|
||||
corruption_levels = [0.1, 0.3, 0.9]
|
||||
num_samples = 100 # Test with multiple samples
|
||||
|
||||
# Test each level
|
||||
for level, expected_corruption in enumerate(corruption_levels):
|
||||
self.curriculum.set_attr_level("corruption_level", level)
|
||||
differences = []
|
||||
|
||||
def test_letter_jumble_iteration():
|
||||
"""Test that iteration respects dataset size"""
|
||||
config = LetterJumbleConfig(size=5, seed=42)
|
||||
dataset = LetterJumbleDataset(config)
|
||||
# Generate multiple problems to measure average corruption
|
||||
for _ in range(num_samples):
|
||||
problem = self.exercise.generate(self.curriculum)
|
||||
metadata = problem["metadata"]["executed_parts"]
|
||||
# Calculate character differences
|
||||
preserve_len = self.curriculum.attributes["preserve_length"].levels[self.curriculum.get_attr_level("preserve_length")]
|
||||
for orig, scrambled in zip(metadata["original_words"], metadata["scrambled_words"]):
|
||||
if len(orig) > preserve_len:
|
||||
diff_count = sum(1 for a, b in zip(orig, scrambled) if a != b)
|
||||
differences.append(diff_count / len(orig))
|
||||
|
||||
items = list(dataset)
|
||||
assert len(items) == config.size
|
||||
# Check average corruption level is reasonable
|
||||
# It's okay if actual corruption is lower than target due to:
|
||||
# 1. Some swaps might cancel out previous swaps
|
||||
# 2. The same characters might be swapped multiple times
|
||||
# 3. The preserve_length attribute prevents some characters from being swapped
|
||||
# 4. For short words, even a few swaps can make them readable
|
||||
if differences:
|
||||
avg_corruption = sum(differences) / len(differences)
|
||||
# Only check that we don't exceed target by too much
|
||||
self.assertLess(avg_corruption, expected_corruption + 0.1,
|
||||
f"Corruption level {avg_corruption:.2f} too high (target: {expected_corruption:.2f})")
|
||||
# And ensure we have some corruption
|
||||
self.assertGreater(avg_corruption, 0.02,
|
||||
f"Corruption level {avg_corruption:.2f} too low (should be above 0.02)")
|
||||
|
||||
# Test multiple iterations yield same items
|
||||
assert items == list(dataset)
|
||||
def test_template_variation(self):
|
||||
"""Test that different templates are used"""
|
||||
templates_seen = set()
|
||||
num_samples = 100
|
||||
|
||||
for _ in range(num_samples):
|
||||
problem = self.exercise.generate(self.curriculum)
|
||||
templates_seen.add(problem["question"].split(":")[0])
|
||||
|
||||
self.assertGreater(len(templates_seen), 1, "Not enough template variation")
|
||||
|
||||
def test_comprehensive_random_evaluation(self):
|
||||
"""Test random evaluation with various configurations and track statistics."""
|
||||
self.rng = random.Random(42) # Fixed seed for reproducibility
|
||||
self.curriculum.rng = self.rng
|
||||
|
||||
# Track statistics
|
||||
word_lengths = defaultdict(int)
|
||||
word_counts = defaultdict(int)
|
||||
corruption_levels = defaultdict(list)
|
||||
consecutive_words_count = 0
|
||||
total_samples = 1000
|
||||
|
||||
# Generate test cases
|
||||
for _ in range(total_samples):
|
||||
# Set random attribute levels
|
||||
for attr in self.curriculum.attributes:
|
||||
max_level = len(self.curriculum.attributes[attr].levels) - 1
|
||||
self.curriculum.set_attr_level(attr, self.rng.randint(0, max_level))
|
||||
|
||||
# Generate and evaluate a random problem
|
||||
problem = self.exercise.generate(self.curriculum)
|
||||
metadata = problem["metadata"]["executed_parts"]
|
||||
original_words = metadata["original_words"]
|
||||
scrambled_words = metadata["scrambled_words"]
|
||||
|
||||
# Track statistics
|
||||
word_counts[len(original_words)] += 1
|
||||
for word in original_words:
|
||||
word_lengths[len(word)] += 1
|
||||
|
||||
# Calculate corruption levels
|
||||
for orig, scrambled in zip(original_words, scrambled_words):
|
||||
preserve_len = self.curriculum.attributes["preserve_length"].levels[self.curriculum.get_attr_level("preserve_length")]
|
||||
if len(orig) > preserve_len:
|
||||
diff_count = sum(1 for a, b in zip(orig, scrambled) if a != b)
|
||||
corruption_levels[len(orig)].append(diff_count / len(orig))
|
||||
|
||||
# Check if words are consecutive in source text
|
||||
if len(original_words) > 1:
|
||||
text = " ".join(self.curriculum.words)
|
||||
phrase = " ".join(original_words)
|
||||
if phrase in text:
|
||||
consecutive_words_count += 1
|
||||
|
||||
# Verify scrambling is valid
|
||||
for orig, scrambled in zip(original_words, scrambled_words):
|
||||
# Check lengths match
|
||||
self.assertEqual(len(orig), len(scrambled))
|
||||
# Check same letters are used
|
||||
self.assertEqual(sorted(orig), sorted(scrambled))
|
||||
|
||||
# Print statistics
|
||||
print("\nWord length distribution:")
|
||||
for length, count in sorted(word_lengths.items()):
|
||||
print(f" Length {length}: {count}")
|
||||
|
||||
print("\nWord count distribution:")
|
||||
for count, freq in sorted(word_counts.items()):
|
||||
print(f" {count} words: {freq}")
|
||||
|
||||
print("\nAverage corruption levels by word length:")
|
||||
for length, levels in sorted(corruption_levels.items()):
|
||||
avg = sum(levels) / len(levels) if levels else 0
|
||||
print(f" Length {length}: {avg:.2f}")
|
||||
|
||||
print(f"\nConsecutive words: {consecutive_words_count}/{total_samples}")
|
||||
|
||||
# Verify statistical properties
|
||||
self.assertTrue(any(length >= 8 for length in word_lengths),
|
||||
"No long words generated")
|
||||
self.assertTrue(any(count >= 3 for count in word_counts.values()),
|
||||
"Not enough variation in word counts")
|
||||
self.assertTrue(consecutive_words_count > 0,
|
||||
"No consecutive words generated")
|
||||
self.assertTrue(consecutive_words_count < total_samples,
|
||||
"Too many consecutive words")
|
||||
|
||||
if __name__ == '__main__':
|
||||
unittest.main()
|
||||
Loading…
Add table
Add a link
Reference in a new issue