A powerful, real-time Chain-of-Thought dataset generator with a beautiful browser UI.

Generate high-quality synthetic training data for LLMs with structured Cold Start + Hot Start reasoning — perfect for creating wop/XXXXXL-style CoT datasets.

• Modern Web UI – Clean, dark-themed interface with live streaming
• True Parallel Generation – Up to 6 concurrent workers using ThreadPoolExecutor
• Real SSE Streaming – Watch tokens appear in real time
• Live Statistics – Progress, speed (chars/sec), ETA, success rate
• Smart Estimator – Accurate remaining time prediction
• Robust Error Handling – Graceful recovery and detailed logging
• JSONL Export – Ready-to-use train.jsonl with optional ASCII cleaning
• One-Click Actions – Copy JSONL, download, refresh, copy streams
• Safe Output – Automatic punctuation normalization & ASCII conversion

pip install requests

Edit the top section of cot_generator.py:

API_URL = "http://127.0.0.1:1234/v1/chat/completions"  # Your local LLM server
MODEL = "qwen/qwen3.5-9b"                               # Change to your model
PORT = 8080
MAX_WORKERS = 6                                         # Adjust based on your hardware

python cot_generator.py

Then open your browser and go to: http://localhost:8080

1. Enter your questions (one per line) in the input box
2. Click ▶ Generate Batch
3. Watch live generation in the stream panel
4. Use the buttons to:
   • Copy JSONL (cleaned & validated)
   • Download train.jsonl
   • Refresh loaded entries

The generator uses a carefully designed system prompt that forces the model to output structured thinking with:

• Cold Start – Initial analysis and context awareness
• Hot Start – Step-by-step reasoning
• <think>...</think> tags (as requested)

You can easily customize the SYSTEM_PROMPT variable in the script.

Each entry follows this structure:

{
  "messages": [
    {
      "role": "system",
      "content": "Enable thinking features: INTUITION, COLD START, HOT START"
    },
    {
      "role": "user",
      "content": "How many R in strawberry"
    },
    {
      "role": "assistant",
      "content": "<think>\n### Cold start\n...\n</think>\n\n**Final Answer**"
    }
  ]
}

• Use a strong reasoning model (Qwen2.5, Llama-3.1/3.3, etc.)
• Keep temperature around 0.7
• Start with 5–20 questions to test
• For very large batches, increase MAX_WORKERS (if your hardware allows)

.
├── cot_generator.py          # Main script (self-contained)
├── train.jsonl               # Generated dataset (auto-created)
└── README.md

Feel free to open issues or PRs! Especially welcome:

• Better UI components
• Support for OpenAI/Anthropic APIs
• Prompt engineering improvements
• Export format options

Made for high-quality synthetic CoT data generation.

Happy dataset building! 🧠