dataset release?
hi there! amazing little model, but would be interesting to apply it to other models like qwen too; do you plan to release the datasets? thanks!
Thanks! I have a plan but haven't decided on the exact release date for our benchmark data yet.
In the meantime, if there's a specific model you need, please let me know. I'll do my best to train the model and upload it to Hugging Face.
Thanks! I have a plan but haven't decided on the exact release date for our benchmark data yet.
In the meantime, if there's a specific model you need, please let me know. I'll do my best to train the model and upload it to Hugging Face.
Could you use Qwen 2.5 1.5b Instruct? I am curious to see how it would perform compared to your llama 3b model.
Well received. We will train Qwen 2.5b first, and let you know once we get the results. Regarding 1.5b models, in our experience, 1b-series models (e.g., Llama-3.2-1b) did not perform well even after preference optimization. Maybe there was some issues regarding training configuration (setups for lr, step size, fewer number of LORA parameters etc), which will need more efforts.
Best,
Well received. We will train Qwen 2.5b first, and let you know once we get the results. Regarding 1.5b models, in our experience, 1b-series models (e.g., Llama-3.2-1b) did not perform well even after preference optimization. Maybe there was some issues regarding training configuration (setups for lr, step size, fewer number of LORA parameters etc), which will need more efforts.
Best,
Interesting, it might be worth trying then Gemma 2b? I have selfish reasons since my intel 11gen laptop gets really slow at processing 700+ tokens on llama 3b.