The model has been fine-tuned using LORA and trained to repeat user messages in ALL CAPS. It took just 11 minutes and 2 epochs (with 4k messages in each) to teach the base 1.6B Stable LM 2 model to follow chat structure and learn the str.upper() behaviour. Trained on RTX 4060 8GB.
!!! Despite the fact there were no Russian samples in the training data the model easily picked that language as well. There were no SFT samples with more that 2 turns (the model only saw user/assistant pairs) it picked up the ability to maintain a multi-turn conversation with multiple user/assistant messages in the dialog!
Trainig code is here.
- Downloads last month
- 11