how to fine tune?
same title
I created an ipynb https://github.com/Deng-Xian-Sheng/Real-technology/blob/main/%E2%80%9CDeepSeek_R1_Distill_Qwen_1_5B_Conversational_ipynb%E2%80%9D%E7%9A%84%E5%89%AF%E6%9C%AC.ipynb
According to my practice, if you try to fine-tune it, your dataset should have a think tag. If not, it will still reason sometimes, but the think tag disappears, which is not conducive to distinguishing answers from reasoning.
Compared with the 7B model, it is more difficult to fine-tune and is more likely to produce repeated sentences. When the repetition penalty is set to 1.5, this problem is alleviated. However, the 7B model has a lower probability of outputting repeated sentences and almost no repetition penalty parameter is needed.
In addition, I only spent 11 minutes on the v100 GPU to fine-tune. The dataset is a 2.6w word json, which contains erotic novels (haha)
If you spend more time to fine-tune, the effect should be improved.
However, for novel writing, a model without think will be better, because in order to fine-tune it, you need a dataset with a think tag, especially a novel, which is not easy to find.
By the way, my 2.6w word dataset can be sold for 30CNY (~=$4). This is a human-written text, and if you want, you can try to translate it into English. I think it's beautiful in any language. Very imaginative fiction. I bought it on the dark web for $100, and it's actually not worth that much.
This ipynb is in Chinese, so you can use Gemini on the top right of Google Colab to translate it cell by cell into English.
This will also help you gradually understand it. Because many people will miss things when they read it directly, and translation is like copying it again.
It is worth noting that this ipynb was originally in English. I slightly modified it (mainly loading the custom dataset on Google Drive in json format and clarifying the dataset format requirements) and translated it into Chinese.
The original ipynb is here: https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Qwen2.5_Coder_(14B)-Conversational.ipynb