Today’s Work
Read a paper translated by an expert: [Paper] LLaMA 2: Open Foundation and Fine-Tuned Chat Models
This Ctrip expert provided both Chinese and English, and translated a bunch of LLM-related papers—respect.
Due to some mysterious force, Hugging Face cannot be git cloned, so I can only use Colab or domestic R&D platforms that support academic acceleration. (Can’t use company internal machines… super painful.)
Deployment
Deployment is actually quite simple, but the hardware requirements are relatively high. Following the Colab deployment method in the project instructions, you can quickly launch a demo.
Of course, the execution steps in Colab can also be run on other R&D platforms that can access Hugging Face.
Next is step two: preparing fine-tuning data.
Data Analysis
As you can see, although the dialogue format is simple, there are quite a few varieties.
So for the first round of training, I’ll remove the narration and fine-tune only on the dialogue. After all, it’s a “dialogue generation” task.
Therefore, for continuous expressions, processing requires checking whether subsequent speakers speak continuously, while also looking for the “」” character to accommodate continuous expressions. Additionally, the inputs and outputs follow a chained A-B, B-C, C-D format. The data volume will be multiplied by 2, but is expected not to exceed 2M.
