Qwen2.5-72B-Instruct-Malaysian

1. Dense LoRA SFT done using DeepSpeed Zero3 HF Trainer.
2. Multipacking variable length 16384 context length, with global batch size of 32, so global total tokens is 524288.
3. All linear layers with rank 256 with alpha multiply by 2.0
4. Liger fused cross entropy.
5. 1e-4 learning rate, 50 warmup, 3 epoch only.

We only upload the best model

Source code

Acknowledgement