A reminder to Microsoft
A reminder to Microsoft:
- The modelโs multilingual capabilities are poor
- The model always loops the same content
- I have done deep zooming and fine-tuning based on the old phi3.5 mini, and felt the real computing and reasoning capabilities (but the larger phi4 does not have this capability)
- Microsoft should try to incorporate more knowledge from its own search engine into the data (rather than being overly "safe (useless)")
- Thank you for open source! I also look forward to Microsoft bringing more shocking open source models to the world!
What do you mean by 2. exactly? @win10 . How do you like this model compared to others, care to explain? Are you using it with Chinese perhaps? My first findings are quite positive, especially for the size.
What do you mean by 2. exactly? @win10 . How do you like this model compared to others, care to explain? Are you using it with Chinese perhaps? My first findings are quite positive, especially for the size.
The ability for Chinese is very weak
And "too safe" is simply not suitable for daily use and literature.
They overly use synthetic data, which is too clean of a distribution, leading to good eval but poor performance. In Phi view, the world is full of rainbows, but in reality, it's full of grammar errors and raw HTML.
Does phi-4 still have the issue of infinite loop where the bot answer is stuck answering only empty strings? I have since always a random issue like that especially when I ask the bot to create a table using markdowns style as usual (at some point it stuck when trying to create the table header / column name).
Bro, for point 2 and 4, I think for Microsoft, it's just because it's open source. We're whoring for nothing...
It's already good enough. In some areas of knowledge, the answers are more detailed than the GPT4O-mini. Thanks open source from M$.
The documentation states that phi-4 is for American English only. Expecting a language model to work with multiple languages would mean a very large dataset.
The documentation states that phi-4 is for American English only. Expecting a language model to work with multiple languages would mean a very large dataset.
I can clearly say:
- The Chinese language of phi3.5 is very good.
- phi4 adds more data training based on the same dataset.
- There is no reason why phi4 should be weaker than phi3.5.
Also, you are right, the problem is that the data set is too small.
- It is clearly stated in the documentation that phi-4 is not intended to support multilingual use. Tokenizers and embedding of Phi4 and Phi3.5 are totally different. It might have added more data, that doesn't mean it supports multilingual.
- Need proof (with temperature, seeds, and proper sampling)
- It might be your cognitive bias (called: confirmation bias). Performance scores prove that it is better than Phi3.5 in the tasks you mentioned.
- No comment on this
- Yes, Phi is a great initiative with open sourcing (unlike non-open-source Llama, Gemma, etc)