3 ms·
I used my 15+ years of IRC logs to fine-tune a GPT-2 model to try to imitate me. My plan was to deploy a bot in my regular IRC channel and see how long it took
by Sohcahtoa82 2y ago
I used my 15+ years of IRC logs to fine-tune a GPT-2 model to try to imitate me. My plan was to deploy a bot in my regular IRC channel and see how long it took for people to notice it was a bot. Any time someone would send a message, it would send the last 10 messages to the LLM, and if the result started with <Sohcahtoa>, then I would send that message to the channel.
Unfortunately, I quickly found GPT-2 isn't nearly good enough. It would generate slightly-coherent yet on-topic nonsense.
Once I overhaul my system, I'll try fine-tuning a 7B model.
- loa_in_ 2y agoI'm pretty sure you could make it work if you expanded the context to hundred or so messages.
- Sohcahtoa82 2y agoWith the 7B parameter models, sure. Not with GPT-2 though. The context window is only 1024 tokens. Even with only 10 messages, if they're long messages, it will exceed the context window.