ChatGpt-nano

nanoChatGpt

a barebones Nanogpt, but finetuned on conversational data

Colab link
all updates in updates.md

how does it work?

This is a fork of Nanogpt, but trained on the data format of a chatbot like chatgpt, with the format inspired by oasst-pythia-12b

<human> ... <endOfText>
<Bot> ... <endOfText>
<human> ... <endOfText>
<Bot> ... <endOfText>
<human> ... <endOfText>
<Bot> ... <endOfText>

Demo

to view its capabilites, simply run the colab, the whole thing should take around ~30 minutes on a t4 instance. If you look at both outputs, Chat has no prompting, but directing the chat in a direction is very helpful

Limitations

I did not make the data dumps/corpuses that make up this data, and can't account for any biases, as the dataset it self is based off the conversations of real people who may or may not have had biases. The model is meant for academic research purposes, and isn't meant for any important or high risk scenarios. Do not follow its advice

whats in the data

• personaChat corpus
• twitter Dumps corpus
• Ubuntu dialog corpus
• Chatbot arena conversations (not for commercial use)
• ParlAI empathetic Chat (not for commercial use)
• Hackernews Dumps corpus
• BabyLM challenge speech data (Unsure, but probably not for commercial use)
• Eleuthier AI arithmetic dataset
• Oasst-1

for commercial purposes, just take the files input1.txt through input36.txt

Features

• Medium Dataset(~630mb), full of a variety of conversations, and a little arithmetic
• can talk to you on a variety of topics, smoothly switch between topics, and often sounds like a real person
• GPT-2-medium 353 million parameters
• Very Fast Inference on GPU
• User to Bot chat
• chat.py has functions similar to openai api stop, removes all content after a certain word
• Option to use google bard to look through and improve conversations

Problems / TODOs

• Math and Logical Reasoning

  • While mostly out of the scope of this model, this is something for future models, the two main options are to detect math and run backend code, or train the model with Chain-of-though math prompting, which is probably more feasible, For logical reasoning, I might try incorporating datasets like garage-bAInd/Open-Platypus, etc

• no memory/recall

  • Though probably out of scope for this model, this is something for future models, With many models, you can ask what you were just talking about or to summarize the conversation above. When that is attempted with this model:

    Human: Dogecoin is cool 
    Bot: indeed, very shibe
    Human: what were we just talking about?
    Bot: me and a friend gave up on mining, but now I can
    

    as we can see, it continues on with a sentence on mining, confirming that it understood the context(GPT2 info) but it does not have the ability to recall. I suspect that has to do with the models data, and that if I were to feed it data like short-context calling and summarization data, it would gain those abilities

<br /><br /><br /> Anyone Who can contribute to the repo, please do so, any and all contributions are welcome, simply add a little to the dataset and expand it dataset would be amazing.

citations

@misc{zheng2023judging,
      title={Judging LLM-as-a-judge with MT-Bench and Chatbot Arena}, 
      author={Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric. P Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica},
      year={2023},
      eprint={2306.05685},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}