nickting/nyt-connections-datasets-raw

NYT Connections Raw Datasets

This repository contains the raw and formatted reasoning data for NYT Connections puzzle solving experiments. These files are the source data used to create the experiment splits in nickting/nyt-connections-experiments.

Overview

This dataset includes three types of puzzle data with AI-generated reasoning:

1. 1.NYT Connections Puzzles - Authentic New York Times puzzles
2. 2.Synthetic Connections Puzzles - Algorithmically generated puzzles
3. 3.Pre-Connections Tasks - Curriculum learning warmup tasks

All data includes both raw reasoning output and formatted conversational format ready for fine-tuning.

Directory Structure

data2/
├── puzzles/                    # Raw puzzle definitions (JSON)
│   ├── connections.json        # NYT puzzle data
│   ├── connections_synthetic.json  # Synthetic puzzle data
│   └── preconn.json            # Pre-Connections tasks
└── reasoning/                  # Generated reasoning data (JSONL)
    ├── structured_nyt_train.jsonl              # NYT structured reasoning (raw)
    ├── structured_nyt_test.jsonl
    ├── structured_nyt_train_formatted.jsonl    # NYT structured reasoning (formatted)
    ├── structured_nyt_test_formatted.jsonl
    ├── structured_synthetic_train.jsonl        # Synthetic structured reasoning (raw)
    ├── structured_synthetic_test.jsonl
    ├── structured_synthetic_train_formatted.jsonl  # Synthetic structured reasoning (formatted)
    ├── structured_synthetic_test_formatted.jsonl
    ├── unstructured_nyt.jsonl                  # NYT unstructured reasoning (raw)
    ├── unstructured_nyt_formatted.jsonl        # NYT unstructured reasoning (formatted)
    ├── unstructured_synthetic.jsonl            # Synthetic unstructured reasoning (raw)
    ├── unstructured_synthetic_formatted.jsonl  # Synthetic unstructured reasoning (formatted)
    ├── structured_preconn_train.jsonl          # Pre-Connections reasoning (raw)
    ├── structured_preconn_test.jsonl
    ├── structured_preconn_train_formatted.jsonl  # Pre-Connections reasoning (formatted)
    └── structured_preconn_test_formatted.jsonl

File Descriptions

Puzzle Files (puzzles/)

connections.json - NYT puzzle definitions

• —Contains 831 authentic NYT Connections puzzles
• —Each puzzle has 16 words grouped into 4 categories
• —Includes metadata: difficulty, category types, publication date

connections_synthetic.json - Synthetic puzzle definitions

• —Contains 200 algorithmically generated puzzles
• —Similar structure to NYT puzzles
• —Generated with categorical constraints

preconn.json - Pre-Connections tasks

• —Contains 800 curriculum learning tasks
• —Simpler than full Connections puzzles
• —Three task types: odd-word-out, find-odd-words, group-formation

Reasoning Files (reasoning/)

Format Types

Structured Format - Reasoning in <think> tags:

{
  "messages": [
    {
      "role": "user",
      "content": "Solve this puzzle: [16 words]"
    },
    {
      "role": "assistant",
      "content": "<think>\nStep 1: Analyze words...\nStep 2: Identify patterns...\n</think>\n\nAnswer: [groups]"
    }
  ]
}

Unstructured Format - Natural narrative reasoning:

{
  "messages": [
    {
      "role": "user",
      "content": "Solve this puzzle: [16 words]"
    },
    {
      "role": "assistant",
      "content": "Looking at these words, I notice... [reasoning]... Therefore, the groups are: [groups]"
    }
  ]
}

Raw vs Formatted

• —Raw files (e.g., structured_nyt_train.jsonl): Initial AI-generated reasoning
• —Formatted files (e.g., structured_nyt_train_formatted.jsonl): Processed into conversational format with proper metadata

Formatted files include additional metadata:

{
  "messages": [...],
  "metadata": {
    "puzzle_id": 95,
    "original_id": 95,
    "permutation": 1,
    "reasoning_length": 2620
  }
}

Data Statistics

NYT Puzzles

• —Total: 831 puzzles
• —Train: 747 puzzles (90%)
• —Test: 84 puzzles (10%)
• —Formats: Structured (3 permutations) + Unstructured (1 permutation)
• —Total entries: ~2,700 structured + ~750 unstructured

Synthetic Puzzles

• —Total: 200 puzzles
• —Train: 180 puzzles (90%)
• —Test: 20 puzzles (10%)
• —Formats: Structured (3 permutations) + Unstructured (1 permutation)
• —Total entries: ~540 structured + ~200 unstructured

Pre-Connections Tasks

• —Total: 800 tasks
• —Train: 720 tasks (90%)
• —Test: 80 tasks (10%)
• —Format: Structured only
• —Total entries: ~800

Data Generation Pipeline

1. 1.Raw Puzzle Data → Generated/collected puzzle definitions
2. 2.Reasoning Generation → AI generates step-by-step solutions
3. 3.gen_reason_struct.py - Structured reasoning
4. 4.gen_reason_unstruct.py - Unstructured reasoning
5. 5.gen_reason_preconn.py - Pre-Connections reasoning
6. 6.Formatting → Convert to conversational format with metadata
7. 7.process_reasoning_format.py - Format Connections data
8. 8.process_preconn_format.py - Format Pre-Connections data
9. 9.Experiment Splits → Create train/validation/test splits
10. 10.prepare_experiments.py → Outputs to nickting/nyt-connections-experiments

Usage

Load Formatted Data for Training

from datasets import load_dataset

# Load NYT structured reasoning
dataset = load_dataset(
    "nickting/nyt-connections-datasets-raw",
    data_files="reasoning/structured_nyt_train_formatted.jsonl"
)

# Load Pre-Connections for curriculum learning
preconn = load_dataset(
    "nickting/nyt-connections-datasets-raw",
    data_files="reasoning/structured_preconn_train_formatted.jsonl"
)

Load Raw Puzzle Definitions

import json

# Load NYT puzzle definitions
with open("puzzles/connections.json") as f:
    nyt_puzzles = json.load(f)

Related Datasets

• —[nickting/nyt-connections-experiments](https://huggingface.co/datasets/nickting/nyt-connections-experiments) - Experiment-ready train/validation/test splits with proper data leakage prevention

Data Quality

• —AI-Generated Reasoning: All reasoning chains are generated by AI, not human-written
• —Validation: Test splits maintained for evaluation
• —Permutations: Multiple word orderings to improve model robustness
• —Format Diversity: Both structured (explicit thinking) and unstructured (narrative) formats

License

MIT License

Citation

@dataset{nyt_connections_raw,
  title={NYT Connections Raw Datasets},
  author={nickting},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/nickting/nyt-connections-datasets-raw}
}

Acknowledgments

• —New York Times for the original Connections puzzle format
• —BigBench for the odd-word-out task inspiration