The Nostradamus experiment

[Sylwy]

Thanks Selwy! If you play several games against Zoe, I will be curious to see the results!

A tournament !

viewtopic.php?p=974384#p974384

Merry Christmas !

[mclane]

Merry Christmas to all. Yes everything now running ….

[David Carteau]

Hi everyone!

For those who are interested, I have just released version v0.3 of the Nostradamus chess engine.

I spent a few days over the summer holidays working on it. My goal was to explore the Transformers architecture and create my own custom model.

I was largely inspired by Andrej Karpathy's excellent tutorial on how LLMs work. Many thanks to him for sharing this material!

https://www.youtube.com/watch?v=kCc8FmEb1nY

The source code is available if you would like to tweak and train your own version of Nostradamus

Regards,

David

[David Carteau]

Hi all,

I'm happy to share a new version (v0.4) of the Nostradamus engine!



This time, I was able to accurately mimic how language models work: the engine takes a 'sentence' as input (a fenstring), converts it into a sequence of 64 tokens (one per square) and then predicts 3 new tokens iteratively, which will encode the best move ('fr' square, 'to' square and promotion, if applicable).

This iterative prediction process drastically reduces the number of illegal moves: once you know which piece is moving, the target square ('to' token) becomes easier to predict

Please do not hesitate to contribute or give feedback to help me improve the model!

David

[Tibono]

Hi David,
refreshing experiment. So eccentric! Thanks for sharing.

Yet I am still using v0.2 thanks to Jim Ablett's Win executable, as I had no clue how to install and run v0.3, despite some instructions you included within the package. I did try, but was definitely puzzled .

[David Carteau]

Hi David,
refreshing experiment. So eccentric! Thanks for sharing.

Yet I am still using v0.2 thanks to Jim Ablett's Win executable, as I had no clue how to install and run v0.3, despite some instructions you included within the package. I did try, but was definitely puzzled .

Hi Tibono,

Thank you for your interest in my little experiment!

If you just want to use the engine, the simplest thing to do is:
- Install Python
- Open a console/terminal and type pip install -U torch==2.7.1 chess==1.11.2 tqdm==4.67.1
- Launch nostradamus.bat

I will try to bundle everything necessary into a ready-to-use executable, but as I have never done this before, it may take me some time: stay tuned!

David

[Tibono]

Thanks a lot, David. Merci infiniment.

I succeeded performing this installation. My first run failed with a runtime error: Numpy is not available.
So I performed this additional command: pip install -U numpy ; resulting in: Successfully installed numpy-2.3.3

My first test game was again itself without openings book, ended with a mate after 57 played moves (so yes, it can give a mate!) but the final stats say: 87/114 random moves (76.32%). Is that the order of magnitude to be expected? I suspect there is an issue with my setup. It plays instantly (whilst v0.2 uses few seconds thinking time before moving). Maybe another python library I should install, even if not preventing the engine to run?

[David Carteau]

I succeeded performing this installation. My first run failed with a runtime error: Numpy is not available.
So I performed this additional command: pip install -U numpy ; resulting in: Successfully installed numpy-2.3.3

My mistake, sorry, I forgot Numpy

My first test game was again itself without openings book, ended with a mate after 57 played moves (so yes, it can give a mate!) but the final stats say: 87/114 random moves (76.32%). Is that the order of magnitude to be expected? I suspect there is an issue with my setup. It plays instantly (whilst v0.2 uses few seconds thinking time before moving). Maybe another python library I should install, even if not preventing the engine to run?

After spending some time trying to understand what is going on, I suspect that there is different behaviour when loading with PyTorch 2.7.1+CPU a network saved with PyTorch 2.7.1+CUDA, resulting in a lot more illegal moves. Perhaps this is because I used BFloat16 for training? I'll investigate.

In the meantime, uninstall PyTorch 2.7.1+CPU and install PyTorch 2.7.1+CUDA using the following commands:
- pip uninstall torch
- pip install torch==2.7.1 --index-url https://download.pytorch.org/whl/cu128

Let me know if this solves your problem!

@Community: help needed there! If anyone has already encountered a CUDA-to-CPU problem, assistance would be very helpful!

Thanks!

David

[David Carteau]

@Community: help needed there! If anyone has already encountered a CUDA-to-CPU problem, assistance would be very helpful!

Ha ha! I suspect a bug in my code! I'm investigating...

[David Carteau]

@Tibono:

The bug has been fixed, but you will need to download the engine again from GitHub. Please confirm that illegal moves have been eliminated

I also tried to build a standalone executable (with PyInstaller): you can download it here (please let me know if it works!).

Regards,

David