open-thought/reasoning-gym
1
0
Fork 0
mirror of https://github.com/open-thought/reasoning-gym.git synced 2026-04-28 17:29:39 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

lint

Browse source
This commit is contained in:
Andreas Koepf 2025-01-30 23:14:32 +01:00
parent 048a165314
commit 5ae329becd
6 changed files with 148 additions and 124 deletions
Download patch file Download diff file Expand all files Collapse all files

67
GALLERY.md
View file

@ -33,6 +33,7 @@ This gallery shows examples from all available datasets using their default conf
- [spell_backward](#spell_backward)
- [sudoku](#sudoku)
- [syllogism](#syllogism)
- [word_ladder](#word_ladder)
- [word_sequence_reversal](#word_sequence_reversal)
- [word_sorting](#word_sorting)
@ -710,19 +711,19 @@ size = 500
Example tasks:
```
Example 1:
Question: How many times does the letter "w" appear in the text: "bed and enters his mechanical dresser Two minutes later the machine deposited him all dressed"?
Answer: 1
Metadata: {'span_length': 15, 'target_letter': 'w', 'span': ['bed', 'and', 'enters', 'his', 'mechanical', 'dresser', 'Two', 'minutes', 'later', 'the', 'machine', 'deposited', 'him', 'all', 'dressed']}
Question: How many times does the letter "o" appear in the text: "bed and enters his mechanical dresser Two minutes later the machine deposited him all dressed"?
Answer: 2
Metadata: {'span_length': 15, 'target_letter': 'o', 'span': ['bed', 'and', 'enters', 'his', 'mechanical', 'dresser', 'Two', 'minutes', 'later', 'the', 'machine', 'deposited', 'him', 'all', 'dressed']}
Example 2:
Question: How many times does the letter "p" appear in the text: "it into a watering place"?
Question: How many times does the letter "c" appear in the text: "it into a watering place"?
Answer: 1
Metadata: {'span_length': 5, 'target_letter': 'p', 'span': ['it', 'into', 'a', 'watering', 'place']}
Metadata: {'span_length': 5, 'target_letter': 'c', 'span': ['it', 'into', 'a', 'watering', 'place']}
Example 3:
Question: How many times does the letter "t" appear in the text: "readable form accessible by the widest array of equipment including outdated"?
Answer: 5
Metadata: {'span_length': 11, 'target_letter': 't', 'span': ['readable', 'form', 'accessible', 'by', 'the', 'widest', 'array', 'of', 'equipment', 'including', 'outdated']}
Question: How many times does the letter "o" appear in the text: "readable form accessible by the widest array of equipment including outdated"?
Answer: 3
Metadata: {'span_length': 11, 'target_letter': 'o', 'span': ['readable', 'form', 'accessible', 'by', 'the', 'widest', 'array', 'of', 'equipment', 'including', 'outdated']}
```
@ -1443,6 +1444,38 @@ Metadata: {'premise1': 'All butterflies are tigers', 'premise2': 'No tigers are
```
### word_ladder
Generates word ladder transformation tasks
Default configuration:
```python
min_word_length = 3
max_word_length = 5
min_chain_length = -1
max_chain_length = -1
seed = 42
size = 500
```
Example tasks:
```
Example 1:
Question: Transform the word 'CEILS' into 'ANIGH' by changing one letter at a time. Each step must create a valid English word (including plurals) and keep the same word length. Show the sequence of words needed.
Answer: CEILS,TEILS,TEINS,THINS,THIGS,THIGH,AHIGH,ANIGH
Metadata: {'start_word': 'CEILS', 'end_word': 'ANIGH', 'word_length': 5, 'chain_length': 8}
Example 2:
Question: Transform the word 'KAW' into 'EFS' by changing one letter at a time. Each step must create a valid English word (including plurals) and keep the same word length. Show the sequence of words needed.
Answer: KAW,KAS,EAS,EFS
Metadata: {'start_word': 'KAW', 'end_word': 'EFS', 'word_length': 3, 'chain_length': 4}
Example 3:
Question: Transform the word 'SAUT' into 'SKER' by changing one letter at a time. Each step must create a valid English word (including plurals) and keep the same word length. Show the sequence of words needed.
Answer: SAUT,SHUT,SHET,SKET,SKER
Metadata: {'start_word': 'SAUT', 'end_word': 'SKER', 'word_length': 4, 'chain_length': 5}
```
### word_sequence_reversal
Generates word sequence reversal tasks from text spans
@ -1491,21 +1524,21 @@ Example tasks:
```
Example 1:
Question: Sort these words in ascending order (using ASCII/Unicode ordering) and return them as a comma-separated list:
due, ever, many, generations
Answer: due, ever, generations, many
Metadata: {'original_words': ['due', 'ever', 'many', 'generations'], 'transformed_words': ['due', 'ever', 'many', 'generations'], 'direction': 'ascending', 'transformation': <TextTransformation.ORIGINAL: 'original'>, 'sorted_words': ['due', 'ever', 'generations', 'many']}
Wolcott, keep, reaching, times
Answer: Wolcott, keep, reaching, times
Metadata: {'original_words': ['Wolcott', 'keep', 'reaching', 'times'], 'transformed_words': ['Wolcott', 'keep', 'reaching', 'times'], 'direction': 'ascending', 'transformation': <TextTransformation.ORIGINAL: 'original'>, 'sorted_words': ['Wolcott', 'keep', 'reaching', 'times']}
Example 2:
Question: Sort these words in descending order (using ASCII/Unicode ordering) and return them as a comma-separated list:
change, 250, young
Answer: young, change, 250
Metadata: {'original_words': ['change', '250', 'young'], 'transformed_words': ['change', '250', 'young'], 'direction': 'descending', 'transformation': <TextTransformation.ORIGINAL: 'original'>, 'sorted_words': ['young', 'change', '250']}
took, critical, condense
Answer: took, critical, condense
Metadata: {'original_words': ['took', 'critical', 'condense'], 'transformed_words': ['took', 'critical', 'condense'], 'direction': 'descending', 'transformation': <TextTransformation.ORIGINAL: 'original'>, 'sorted_words': ['took', 'critical', 'condense']}
Example 3:
Question: Sort these words in ascending order (using ASCII/Unicode ordering) and return them as a comma-separated list:
industry, elementary, traverse, stepped, meals, rub, resultant, etheric, irritation
Answer: elementary, etheric, industry, irritation, meals, resultant, rub, stepped, traverse
Metadata: {'original_words': ['industry', 'elementary', 'traverse', 'stepped', 'meals', 'rub', 'resultant', 'etheric', 'irritation'], 'transformed_words': ['industry', 'elementary', 'traverse', 'stepped', 'meals', 'rub', 'resultant', 'etheric', 'irritation'], 'direction': 'ascending', 'transformation': <TextTransformation.ORIGINAL: 'original'>, 'sorted_words': ['elementary', 'etheric', 'industry', 'irritation', 'meals', 'resultant', 'rub', 'stepped', 'traverse']}
apartment, yellow, Just, pleasure, collapse, different, purchasers, taking, opening
Answer: Just, apartment, collapse, different, opening, pleasure, purchasers, taking, yellow
Metadata: {'original_words': ['apartment', 'yellow', 'Just', 'pleasure', 'collapse', 'different', 'purchasers', 'taking', 'opening'], 'transformed_words': ['apartment', 'yellow', 'Just', 'pleasure', 'collapse', 'different', 'purchasers', 'taking', 'opening'], 'direction': 'ascending', 'transformation': <TextTransformation.ORIGINAL: 'original'>, 'sorted_words': ['Just', 'apartment', 'collapse', 'different', 'opening', 'pleasure', 'purchasers', 'taking', 'yellow']}
```

91
examples/generate_word_ladder_examples.py
View file

@ -1,48 +1,50 @@
# generates dataset of word ladder examples, and then generates simulated chain of thought reasoning for each example
import reasoning_gym
from openai import OpenAI
import os
from openai import OpenAI
import reasoning_gym
# Configuration for the dataset
config = {
'dataset_name': 'word_ladder',
'dataset_config': {
'min_word_length': 5,
'max_word_length': 5,
'min_chain_length':3, # set to -1 for shortest possible path, increase to generate more examples
'max_chain_length':5,
'size': 1, # Generate a small dataset for demonstration
}
"dataset_name": "word_ladder",
"dataset_config": {
"min_word_length": 5,
"max_word_length": 5,
"min_chain_length": 3, # set to -1 for shortest possible path, increase to generate more examples
"max_chain_length": 5,
"size": 1, # Generate a small dataset for demonstration
},
}
system_prompt = """Word Ladder puzzles involve transforming a start word into an end word.
You are allowed to change only one letter a time and you must keep the number of letters constant.
Each time you change one letter the word in the chain must be forming one that's valid in English.
Plurals are allowed, but not proper nouns.
system_prompt = """Word Ladder puzzles involve transforming a start word into an end word.
You are allowed to change only one letter a time and you must keep the number of letters constant.
Each time you change one letter the word in the chain must be forming one that's valid in English.
Plurals are allowed, but not proper nouns.
Given a start and an end word, generate a detailed step-by-step chain of thought reasoning of the transformation process.
You will be given the word ladder question, as well as the correct solution path. So you don't need to solve the problem, you have the solution.
Your task is to provide a perfectly simulated chain of thought reasoning exactly in the style of the example below,
including considering multiple possibilities, validating words, and showing the final path.
Your task is to provide a perfectly simulated chain of thought reasoning exactly in the style of the example below,
including considering multiple possibilities, validating words, and showing the final path.
Use casual, verbose thinking with markdown formatting.
Example chain of thought reasoning:
Example chain of thought reasoning:
{sample_CoT}
"""
# sample CoT for demonstration, extracted from real inference data from deepseek r1
sample_CoT = """
Word Ladder puzzles involve transforming a start word into an end word.
You are allowed to change only one letter a time and you must keep the number of letters constant.
Each time you change one letter the word in the chain must be forming one that's valid in English.
Plurals are allowed, but not proper nouns.
sample_CoT = """
Word Ladder puzzles involve transforming a start word into an end word.
You are allowed to change only one letter a time and you must keep the number of letters constant.
Each time you change one letter the word in the chain must be forming one that's valid in English.
Plurals are allowed, but not proper nouns.
Given a start and an end word, generate a detailed step-by-step chain of thought reasoning of the transformation process.
You will be given the word ladder question, as well as the correct solution path. So you don't need to solve the problem, you have the solution already, which should be your 'cheat sheet' to refer to as you go along.
Your task is to provide a perfectly simulated chain of thought reasoning exactly in the style of the example below,
including considering multiple possibilities, validating words, and showing the final path.
Your task is to provide a perfectly simulated chain of thought reasoning exactly in the style of the example below,
including considering multiple possibilities, validating words, and showing the final path.
Use casual, verbose thinking with markdown formatting.
VERY IMPORTANT ***NEVER MENTION YOU HAVE BEEN PROVIDED WITH THE ANSWER, YOU NEED TO WRITE ONLY AS IF YOU'RE SOLVING THE QUESTION STEP BY STEP ***
Example chain of thought reasoning:
Example chain of thought reasoning:
<think>
Okay, so I need to figure out how to transform the word "CARK" into "ZIPS" by changing one letter at a time, making sure each intermediate step is a valid English word. Let me start by breaking down the problem.
@ -177,44 +179,37 @@ Yes, that's four steps, each changing one letter, all words are valid. So that s
"""
client = OpenAI(
base_url="https://openrouter.ai/api/v1", # openrouter used in this example
api_key=os.environ["OPENROUTER_API_KEY"], # set your key in environment variable
base_url="https://openrouter.ai/api/v1", # openrouter used in this example
api_key=os.environ["OPENROUTER_API_KEY"], # set your key in environment variable
)
def generate_cot(question: str, answer: str) -> str:
"""Generate chain of thought reasoning for word ladder"""
prompt = f"""The question is {question}. The correct solution is {answer}.
prompt = f"""The question is {question}. The correct solution is {answer}.
Provide the verbose chain of thought reasoning to transform the start word into the end word exactly in the style and length required."""
completion = client.chat.completions.create(
model="microsoft/phi-4", # choose model
messages=[
{
"role": "system",
"content": system_prompt
},
{
"role": "user",
"content": prompt
}
],
model="microsoft/phi-4", # choose model
messages=[{"role": "system", "content": system_prompt}, {"role": "user", "content": prompt}],
temperature=0.6,
max_tokens=10000
max_tokens=10000,
)
return completion.choices[0].message.content
# Create the word ladder dataset
dataset = reasoning_gym.create_dataset(config['dataset_name'], **config['dataset_config'])
dataset = reasoning_gym.create_dataset(config["dataset_name"], **config["dataset_config"])
print(f"Generated {len(dataset)} examples, moving on to generate CoT reasoning...")
# Generate and print examples with CoT
for item in dataset:
# Generate CoT reasoning demo
item['reasoning'] = generate_cot(item['question'],item['answer'])
item["reasoning"] = generate_cot(item["question"], item["answer"])
print("\n--- Example ---")
print("Question:", item['question'])
print("Answer:", item['answer'])
print("Question:", item["question"])
print("Answer:", item["answer"])
print("\nChain of Thought:")
print(item['reasoning'])
print("\nMetadata:", item['metadata'])
print(item["reasoning"])
print("\nMetadata:", item["metadata"])

2
reasoning_gym/algorithmic/__init__.py
View file

@ -14,9 +14,9 @@ from .number_filtering import NumberFilteringConfig, NumberFilteringDataset
from .number_sorting import NumberSortingConfig, NumberSortingDataset
from .sentence_reordering import SentenceReorderingConfig, SentenceReorderingDataset
from .spell_backward import SpellBackwardConfig, SpellBackwardDataset
from .word_ladder import WordLadderConfig, WordLadderDataset
from .word_sequence_reversal import WordSequenceReversalConfig, WordSequenceReversalDataset
from .word_sorting import TextTransformation, WordSortingConfig, WordSortingDataset
from .word_ladder import WordLadderConfig, WordLadderDataset
__all__ = [
"SpellBackwardConfig",

97
reasoning_gym/algorithmic/word_ladder.py
View file

@ -1,46 +1,51 @@
"""Word ladder task generator"""
from collections import deque
from dataclasses import dataclass
from random import Random
from typing import List, Optional, Set, Dict, Tuple
from collections import deque
from typing import Dict, List, Optional, Set, Tuple
from reasoning_gym.data import read_data_file
from ..factory import ProceduralDataset, register_dataset
@dataclass
class WordLadderConfig:
"""Configuration for word ladder task generation"""
min_word_length: int = 3 # Minimum word length
max_word_length: int = 5 # Maximum word length
min_chain_length: int = -1 # Set to -1 for shortest path or a minimum of 3
max_chain_length: int = -1 # Set to -1 for shortest path or a max
min_word_length: int = 3 # Minimum word length
max_word_length: int = 5 # Maximum word length
min_chain_length: int = -1 # Set to -1 for shortest path or a minimum of 3
max_chain_length: int = -1 # Set to -1 for shortest path or a max
seed: Optional[int] = None
size: int = 500 # Virtual dataset size
size: int = 500 # Virtual dataset size
def validate(self) -> None:
"""Validate configuration parameters"""
assert self.min_word_length > 2, "min_word_length must be 3"
assert self.max_word_length >= self.min_word_length, "max_word_length must be >= min_word_length"
assert self.max_word_length <= 5, "max_word_length must be 5"
# Modified validation logic
if self.min_chain_length == -1:
if self.max_chain_length != -1:
assert self.max_chain_length >= 3, "When min_chain_length=-1 (shortest path), max_chain_length must be -1 or >=3"
assert (
self.max_chain_length >= 3
), "When min_chain_length=-1 (shortest path), max_chain_length must be -1 or >=3"
elif self.max_chain_length == -1:
raise AssertionError("max_chain_length cannot be -1 unless min_chain_length is also -1")
else:
assert self.min_chain_length >= 3, "min_chain_length must be 3 or -1"
assert self.max_chain_length >= self.min_chain_length, "max_chain_length must be >= min_chain_length"
class WordLadderDataset(ProceduralDataset):
"""Generates word ladder transformation tasks"""
def __init__(self, config: WordLadderConfig):
super().__init__(config=config, seed=config.seed, size=config.size)
# Load words from CSV file
self.word_sets = self._load_words_from_csv()
@ -48,36 +53,37 @@ class WordLadderDataset(ProceduralDataset):
"""Load words from CSV file organized by length"""
import csv
from io import StringIO
word_sets = {}
try:
# Get CSV content as string
csv_content = read_data_file("words.csv")
# Use StringIO to create a file-like object from the string
csv_file = StringIO(csv_content)
reader = csv.DictReader(csv_file)
for row in reader:
# Process each word length column
for length in range(3, 6):
col_name = f'{length}_letter'
word = row.get(col_name, '')
col_name = f"{length}_letter"
word = row.get(col_name, "")
if not word: # Skip empty entries
continue
if self.config.min_word_length <= length <= self.config.max_word_length:
word_sets.setdefault(length, set()).add(word.upper())
except Exception as e:
raise RuntimeError(f"Error processing words.csv content: {e}") from e
# Validate we have words for each length
for length in range(self.config.min_word_length, self.config.max_word_length + 1):
if length not in word_sets or not word_sets[length]:
raise ValueError(f"No valid words found for length {length}")
return word_sets
def _differs_by_one(self, word1: str, word2: str) -> bool:
@ -96,16 +102,16 @@ class WordLadderDataset(ProceduralDataset):
"""Find path between start and end words that meets length requirements"""
if start == end:
return [start]
# First find shortest path length
shortest_path = self._bfs_shortest_path(start, end, word_set)
if not shortest_path:
return None
min_length = self.config.min_chain_length
if len(shortest_path) > min_length:
return shortest_path # Shortest path is already longer than required
# Now look for longer paths using DFS with depth constraint
return self._dfs_with_depth(start, end, word_set, min_length)
@ -113,12 +119,12 @@ class WordLadderDataset(ProceduralDataset):
"""BFS implementation to find shortest path"""
queue = deque([(start, [start])])
visited = {start}
while queue:
current, path = queue.popleft()
if current == end:
return path
for neighbor in self._get_neighbors(current, word_set):
if neighbor not in visited:
visited.add(neighbor)
@ -128,62 +134,62 @@ class WordLadderDataset(ProceduralDataset):
def _dfs_with_depth(self, start: str, end: str, word_set: Set[str], target_length: int) -> Optional[List[str]]:
"""DFS implementation looking for paths of exact length"""
stack = [(start, [start], set([start]))]
while stack:
current, path, visited = stack.pop()
if len(path) == target_length:
if current == end:
return path
continue
if len(path) > target_length:
continue
# Explore neighbors in random order to find different paths
neighbors = list(self._get_neighbors(current, word_set))
Random().shuffle(neighbors)
for neighbor in neighbors:
if neighbor not in visited:
new_visited = set(visited)
new_visited.add(neighbor)
stack.append((neighbor, path + [neighbor], new_visited))
return None
def _get_neighbors(self, word: str, word_set: Set[str]) -> Set[str]:
"""Get all valid neighbors that differ by one letter"""
neighbors = set()
word_chars = list(word)
for i in range(len(word_chars)):
original = word_chars[i]
for c in 'ABCDEFGHIJKLMNOPQRSTUVWXYZ':
for c in "ABCDEFGHIJKLMNOPQRSTUVWXYZ":
if c == original:
continue
word_chars[i] = c
new_word = ''.join(word_chars)
new_word = "".join(word_chars)
if new_word in word_set:
neighbors.add(new_word)
word_chars[i] = original
return neighbors
def _generate_word_pair(self, rng: Random, length: int) -> Tuple[str, str, List[str]]:
"""Generate valid start/end words with solution path"""
word_set = self.word_sets[length]
max_attempts = 500
for _ in range(max_attempts):
start, end = rng.sample(sorted(word_set), 2)
path = self._find_path(start, end, word_set)
if path and (
(self.config.min_chain_length == -1 and self.config.max_chain_length == -1) or
(self.config.min_chain_length <= len(path) <= self.config.max_chain_length)
(self.config.min_chain_length == -1 and self.config.max_chain_length == -1)
or (self.config.min_chain_length <= len(path) <= self.config.max_chain_length)
):
return start, end, path
raise RuntimeError(f"Failed to find valid pair for length {length} after {max_attempts} attempts")
def __getitem__(self, idx: int) -> dict:
@ -191,17 +197,12 @@ class WordLadderDataset(ProceduralDataset):
rng = Random(self.seed + idx)
length = rng.randint(self.config.min_word_length, self.config.max_word_length)
start, end, path = self._generate_word_pair(rng, length)
return {
"question": f"Transform the word '{start}' into '{end}' by changing one letter at a time. Each step must create a valid English word (including plurals) and keep the same word length. Show the sequence of words needed.",
"answer": ",".join(path),
"metadata": {
"start_word": start,
"end_word": end,
"word_length": length,
"chain_length": len(path)
}
"metadata": {"start_word": start, "end_word": end, "word_length": length, "chain_length": len(path)},
}
register_dataset("word_ladder", WordLadderDataset, WordLadderConfig)
register_dataset("word_ladder", WordLadderDataset, WordLadderConfig)

2
reasoning_gym/data/words.csv
View file

@ -12960,4 +12960,4 @@ ZZZ,ELLS,BOSKY
,,ZYGAL
,,ZYGON
,,ZYMES
,,ZYMIC
,,ZYMIC

1 3_letter 4_letter 5_letter
12960 ZYGAL
12961 ZYGON
12962 ZYMES
12963 ZYMIC

13
tests/test_word_ladder.py
View file

@ -44,12 +44,7 @@ def test_word_ladder_dataset_deterministic():
def test_word_ladder_dataset_items():
"""Test basic properties of generated items"""
config = WordLadderConfig(
min_word_length=3,
max_word_length=5,
min_chain_length=3,
max_chain_length=5,
size=10,
seed=42
min_word_length=3, max_word_length=5, min_chain_length=3, max_chain_length=5, size=10, seed=42
)
dataset = WordLadderDataset(config)
@ -76,7 +71,7 @@ def test_word_ladder_dataset_items():
# Verify solution chain from answer
solution_chain = item["answer"].split(",")
# Handle chain length validation based on whether it's shortest path (-1) or specified length
if metadata["chain_length"] == -1:
# For shortest path, just ensure it's a valid path (we can't predict exact length)
@ -85,7 +80,7 @@ def test_word_ladder_dataset_items():
# For specified length, ensure it matches config constraints
assert config.min_chain_length <= len(solution_chain) <= config.max_chain_length
assert len(solution_chain) == metadata["chain_length"]
assert solution_chain[0] == metadata["start_word"]
assert solution_chain[-1] == metadata["end_word"]
assert all(len(word) == word_length for word in solution_chain)
@ -144,4 +139,4 @@ def test_word_ladder_find_path():
if __name__ == "__main__":
pytest.main([__file__])
pytest.main([__file__])