Trained on 100k dumped messages from the 'chan' todd proxy. I could not dedupe the dataset but it has had serious effect on the llama7b I used. Calls me master a whole bunch more now.
Content isn't SFW so be aware. Trained in 4-bit for 3 epochs, I think it overfit and really needed just 2.
Tested in 4-bit and FP16 on plain HF llama-7b, maybe it works on derivative models of the same beaks.
V2 version was trained at a higher rank and logner context (512) on only unique data with ALLMs and "content warning" statements removed. It is much stronger.