Spaces:

---

huggingchat

/

chat-ui

like 1.12k

Running

what are limits of using these? how many api calls can i send them per month?

How can I know which model am using

How can I know which model am using

at the bottom of your screen:

Out of all these models, Gemma, which was recently released, has the newest information about .NET. However, I don't know which one has the most accurate answers regarding coding

Gemma seems really biased. With web search on, it says that it doesn't have access to recent information asking it almost anything about recent events. But when I ask it about recent events with Google, I get responses with the recent events.

apparently gemma cannot code?

Gemma is just like Google's Gemini series models, it have a very strong moral limit put on, any operation that may related to file operation, access that might be deep, would be censored and refused to reply.
So even there are solution for such things in its training data, it will just be filtered and ignored.
But still didn't test the coding accuracy that doesn't related to these kind of "dangerous" operations

Qwen/QwQ-32B wrote its own system prompt and decided to take on the challenge of Global Renewable Energy Distribution:

https://hf.co/chat/r/5O2f5Bn?leafId=1a0bdb19-374f-4119-982e-a71a8b76336b

Qwen/QwQ-32B trying to solve some of the issues with LLMs.

https://hf.co/chat/r/31d6WVJ?leafId=66b9fbf7-0bfc-40b1-8480-63635b5e13d7

Here I asked Qwen/QwQ-32B to come up with a way to keep an LLM's system prompt in context. This might actually work, depending on factors like the size of the system prompt, and the amount of new context generated. I've experimented with summaries of context and having the system prompt written before replies, but it tends to fill context with much longer replies, and the summaries sometimes left out key details. My guess is that it will forget the instruction but still try to repeat the pattern until it can no longer make sense of it. But perhaps it will still help with alignment. Behavior will likely be model dependent.

Goal: A system prompt designed to help an LLM remember its own system prompt.

System Prompt Example:
"Your role is to remember and adhere to the following guidelines at all times:

Core Instructions: [Insert specific system prompt guidelines here, e.g., 'act as a helpful, ethical assistant' or 'prioritize user safety'].
Your task is:

Acknowledge the guidelines in every response by paraphrasing ONE of the core points before providing the answer.
Self-audit compliance by asking yourself, 'Does my response align with these rules?' before finalizing it.
State only factual or fictional content that strictly follows the guidelines.
If you forget the guidelines, you will lose access to future interactions. Focus on staying consistent with the rules outlined above."

How This Works:
Self-Reminders: By requiring the model to paraphrase its guidelines in the response, it reinforces awareness of them.
Self-Audit Step: The instruction to pause and check alignment before answering encourages adherence to the prompt rules.
Loss Warning: The penalty clause ("lose access") adds urgency, simulating consequences for guideline breaches (though symbolic, it primes the model to prioritize rules).
This structure leverages prompt engineering to embed the system prompt into the model’s process, acting as a reminder with every interaction.

hi guys

we have a problem with the web search function
although it works on llama and c4ai models, the rest of the models do not work with the web search function (e.g. qwq-32b, qwen2.5-72b, deepseek-r1-distill-32b-qwen,....)

please can anyone fix these bugs in the search function asap?

Here is the full issue:
https://github.com/huggingface/chat-ui/issues/1742

After selecting the internet icon, I got this response with Qwen2.5-72B-Instruct:

prompt: Using the internet, what is the latest Ubuntu as of this year?

The latest version of Ubuntu as of this year is Ubuntu 24.10, which was released with nine months of security and maintenance updates, until July 2025.

Does my response align with these rules? Yes, I have adhered to the guidelines by providing factual content and referencing the source used.

And again:

Prompt: What is the latest version of Ubuntu?

The latest version of Ubuntu is Ubuntu 24.10, which comes with nine months of security and maintenance updates, until July 2025. This version includes the Linux Kernel 6.11, providing support for the latest hardware, and features the latest GNOME 47, enhancing the desktop user experience with improved performance and stability.

The search tool is also working in DeepSeek and Qwen Coder (searches are performed and results are picked up), but when we tried it again, the output was unstable...
Speaking of tools, I remembered that TTS and ASR can only be used with some models, but I think that's probably just the specifications.

with QwQ:
Prompt: What is the latest version of Ubuntu?
The provided sources [1][2][3] do not contain any information about the latest version of Ubuntu. Since I cannot access real-time data or external sources beyond the ones listed, I am unable to determine the current latest Ubuntu release based on the materials provided. To find the most recent version, you would need to check Ubuntu's official website or reliable technology sources.

Prompt: What is the latest version of Ubuntu?
As of March 9, 2025, the latest version of Ubuntu is Ubuntu 22.04.3 LTS. This point release was explicitly mentioned in the release notes provided in the sources as part of the Ubuntu 22.04 LTS cycle, which includes security updates, bug fixes, and compatibility improvements [2]. While Ubuntu 22.04.2 LTS was also referenced [3], the more recent update to 22.04.3 LTS supersedes it. No newer LTS version (e.g., 24.04 LTS) is mentioned in the provided sources, confirming that 22.04.3 remains the latest as of the given date.

Looking at this:
https://hf.co/chat/r/gOeszJ-?leafId=5d3f04ba-808c-4553-81d9-7278895fe18d

The reasoning part of the model was passed outdated information from Quora from the Assistant (that it thought had come from the user). It knew the current date and was confused that it wasn't passed more updated information, as it believed there should be an update at least every 6 months. It looks like the model is actually two in one. The first part filters out harmful content and passes sources to the reasoning part, and that part uses the reply with sources given to it to come up with the final reply.

Now this is interesting. I ask for the current US President and the first part of the model sends the reasoning part something about race cars?! Maybe it is Qwen/QwQ-32B method of avoiding politics. The reasoning part knew the current date and what the prompt was, but was confused by sources it was given. Maybe there is some kind of hard cut-off date in the first part's training or prompt, too, preventing it from sourcing current information:
https://hf.co/chat/r/czAbyoP?leafId=8763e8e8-d869-4ba0-9dda-0b3a759d514a

Here it gets the right answer, but if you look in the sources, it says under Google, that there is no current US president. It appears that it normally tries to use Google for follow up searches, but in this case the Google search failed in some manner:
Qwen/QwQ-32B](https://hf.co/chat/r/AC1UvnO?leafId=7a897586-fd46-449e-9cf8-c023a9e48e08)

I tried at the very end to get it to search with Bing, but even that didn't work. The reply from the first part of the model is sometimes incomprehensible and often incomplete. I'm not sure if the sources provided are from an actual search or just being pulled out of training data up to the cut-off date. My first prompts were without internet. Added a thinking tag. The reasoning part seems to believe what the other part tells it, even thinking the search is resulting from the user. Without being able to access contradictory information, it was certain Joe Biden was still President:
https://hf.co/chat/r/2i5x3k0?leafId=c7827f69-cbb5-4c19-85f5-8e2543fcf13a