hyp1823/DeepSeek-R1-lottery

Since the model is public and somebody will download the model, it is necessary for me to write the following explanation:

The model is intended to be used by my own Taiwanese lottery prediction project, specifically the 539 lottery game in Taiwan.

The Taiwanese lottery history data can be crawled using https://github.com/stu01509/TaiwanLotteryCrawler

And the naive thought that came up to me about 2~3 years ago is that lottery numbers are seqeunces, and lottery prediction using past history numbers can be seen as sequence to sequence prediction problem.

The LLMs that are trending in recent years are suitable to be used to finetune on the numbers sequences.

The way I did is to convert the numbers to some fixed random chosed words (using words in vocab.txt, convert 1 to romanesque for example), and generate suitable data like this (for a 539 prediction, using a last numbers to predict next numbers):

                 question:  cautious sanctions marino scroll theologian
                 response:  jain surveyor searches roaring connell

of course the output needs to be convert back to numbers...

Since the model is finetuned, the model can be seen as a lottery prediction model but contains language meaning/distribution.

Also, one thing to notice is that I believe that the lottery game might be exploitable when the game history now is relatively short, and if the history goes longer and loger such that every combination of sequences happens, then it's not suitable to train/finetune on all of the data, one might need to consider the data selection problem, maybe using just the recent data for training, or do some statistics and dynamically selects the data with importance.

I didn't do any evaluation on the model, I just finetuned it, I didn't even notice the performance metrics (the loss ?). So I'm not quite sure about the performance on the history data, but I don't think it's that important because past performance doesn't mean future performance in this random problem.

I use the post here https://www.datacamp.com/tutorial/fine-tuning-deepseek-r1-reasoning-model and my own data conversion to finetune this model, it actually use kaggle because there are free gpu hours.

hope this helps !!!

Uploaded model

• Developed by: hyp1823
• License: apache-2.0
• Finetuned from model : unsloth/deepseek-r1-distill-llama-8b-unsloth-bnb-4bit

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.