[MODELS] Discussion

#372
by victor HF staff - opened
Hugging Chat org
โ€ข
edited 20 days ago

Here we can discuss about HuggingChat available models.

image.png

victor pinned discussion

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:
image.png

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

This comment has been hidden

is it possible to know what parameters this models are running ?

Hugging Chat org

is it possible to know what parameters this models are running ?

It's all here! https://github.com/huggingface/chat-ui/blob/main/.env.template

is it possible to know what parameters this models are running ?

It's all here! https://github.com/huggingface/chat-ui/blob/main/.env.template

thanks this is super useful OWO

What happened to Falcon? It was my favorite. :(

Hugging Chat org
โ€ข
edited Feb 27

@SAMMdev Falcon was too costly to run at scale (for now), we might put back a more optimized version in the future

I would like to use "mistralai/Mixtral-8x7B-Instruct-v0.1";
Please could tell me what is the precision of the model behind the chat? Thanks

This comment has been hidden

@SAMMdev Falcon was too costly to run at scale (for now), we might put back a more optimized version in the future

What if we use Falcon 70B?

smaug 72B would be a great addition

Iโ€™m unable to get output from CodeLlama

I'm also voting for Samsung 72B. We already have the two Llama 70B models on here soo to me it seems reasonable to integrate this one as well.

This is probably not going to happen, but xai-org/grok-1 would be insane to have here

IYH Why is the title of most chats (on the left panels roster) "๐Ÿค– Hello! I am a language model AI assistant."?

This implies that the system prompt of my assistants is not the fundamental prompt, but there is an inbuilt base prompt that is run before my system prompt .. is this correct roughly and if so how do I change this base prompt for Mistral LLM ?

hgchattitles.PNG

Could you consider DBRX Instruct and Command-R? The official space for DBRX Instruct is too limited (it only allows for a 5-turn conversation) and there is no space for Command-R.

IYH thank you for your advice. Apologies I have no idea what the concepts mean or what to do "Could you consider DBRX Instruct and Command-R? The official space for DBRX Instruct is too limited (it only allows for a 5-turn conversation) and there is no space for Command-R." (fwiw I prompted mistral about it and it did not know either.)
Would you kindly elaborate (or point me towards a resource that explains this

IYH thank you for your advice. Apologies I have no idea what the concepts mean or what to do "Could you consider DBRX Instruct and Command-R? The official space for DBRX Instruct is too limited (it only allows for a 5-turn conversation) and there is no space for Command-R." (fwiw I prompted mistral about it and it did not know either.)
Would you kindly elaborate (or point me towards a resource that explains this

Huggingface will notify you when someone posts in a discussion you've commented on, even if they didn't directly reply to you. I was suggesting two new models, unrelated to your question.

Which model is better to use?
How to know difference of them?

IYH Why is the title of most chats (on the left panels roster) "๐Ÿค– Hello! I am a language model AI assistant."?

This implies that the system prompt of my assistants is not the fundamental prompt, but there is an inbuilt base prompt that is run before my system prompt .. is this correct roughly and if so how do I change this base prompt for Mistral LLM ?

hgchattitles.PNG

@DYB5784 HF Chat has a Mistral 7B model setup with system prompt for the task of summarizing the first chat prompt/msg into a title for the chat history log, so unless one explicitly addresses that in the first msg it is what it is ig. and we can always rename it. Still, i think it would have been awesome if we could customize the naming style/prompt it ourselves.

Is openchat/openchat-3.5-0106 coming back? Was it removed to be upgraded?

Is openchat/openchat-3.5-0106 coming back? Was it removed to be upgraded?

It looks like they also removed the Meta models :(

Hope they add command r instead of bringing those back tbh.

Hope they add command r instead of bringing those back tbh.

What is command r? I'm a newb.

Hope they add command r instead of bringing those back tbh.

What is command r? I'm a newb.

Command-r+ is a new LLM from Cohere that overtook GPT-4 on the openllm leaderboard.

Hugging Chat org

hey!

On HuggingChat we aim to always propose a small selection of models which will evolve over time as the field of ML progresses forward ๐Ÿ”ฅ

Stay tuned!

On Hugging Chat we aim to always propose a small selection of models which will evolve over time as the field of ML progresses forward ๐Ÿ”ฅ
Stay tuned!

Yup, small models are better and lighter (Cost friendly) + now Hugging chat ai has internet access so small models like (Mixtrail, Nous hermes, etc.) can even performs very better in many areas then many 70b models, and
We are happy to see what's coming next ๐Ÿ”ฅ๐Ÿ”ฅ.

Hope they add command r instead of bringing those back tbh.

The Meta ones felt misaligned and gave a lot of refusals. The 70b code one would lecture and moralize even with nothing bad in the prompt.

I hope LLaMA 3 isn't as misaligned mess.

The Meta ones felt misaligned and gave a lot of refusals. The 70b code one would lecture and moralize even with nothing bad in the prompt.

This is because they do not do fine tuning, Manytime fully Finetuned model of Llama 7b is better than no fine tuned llama 70b

Hugging Chat org
โ€ข
edited 20 days ago

Cohere Command R+ is now on HuggingChat!

image.png

@Victor Thank you for the new model! but if possible, i think a slight warning/notification should be very helpful to us about which model will be taken down!
Goodbye, OpenChat! it was really good for 7B!

@Victor Thank you for the new model! but if possible, i think a slight warning/notification should be very helpful to us about which model will be taken down!
Goodbye, OpenChat! it was really good for 7B!

Agree with 1st point

@Victor Thank you for the new model! but if possible, i think a slight warning/notification should be very helpful to us about which model will be taken down!
Goodbye, OpenChat! it was really good for 7B!

Agree

Disagree...

Cohere Command R+ is now on HuggingChat!

image.png

... Hey Victor, if you're gonna surprise us with new models like this, then you can remove anything you want without notify anyone, not even Clement, xd.

But jokes aside, this is just great, if your adding/removal policy keeps like this, in 3 months we will have hugging-face assistants for everything, long context, coding, reasoning/creativity, etcetera.

Thanks a lot!!!

P.D.: I was expecting just Command R, but having plus with all the HF interface means that I will be able to make a lot of assistants that on the past only worked decently as GPTs with GPT4.

@Victor Thank you for the new model! but if possible, i think a slight warning/notification should be very helpful to us about which model will be taken down!
Goodbye, OpenChat! it was really good for 7B!

Agree

Disagree...

why bro

you can remove anything you want without notify anyone

@Ironmole you're literally ok with leaving all the active chats abandoned, aren't you? what can we say here? but lots of users will be kinda saddened if active/hanging chats are suddenly no longer continuable. (i know they usually take down the models with the least traffic, so, that's how it is ig)
and it seems all the other assistants have been migrated to mistralai/Mixtral-8x7B-Instruct-v0.1.

Hugging Chat org

Yes we migrated all assistants with deprecated models to the default model, which at the time was Mixtral 8x7B!

command r + is really good

Iโ€™m worried about that itโ€™s not gonna be free forever, like donโ€™t get me wrong I have FULL faith in the hugging chat team, itโ€™s just this in my eyes itโ€™s a perfect replacement to ChatGPT. So I just need some reassurance itโ€™ll stay free

Iโ€™m worried about that itโ€™s not gonna be free forever, like donโ€™t get me wrong I have FULL faith in the hugging chat team, itโ€™s just this in my eyes itโ€™s a perfect replacement to ChatGPT. So I just need some reassurance itโ€™ll stay free

I think that It'll stay free.
But if they have budget issue then.
They can integrate ads to make it free forever.
and also introduce premium features (Like some premium model only use by premium or Badge to pro, etc.)

Please leave that Command R plus unquantized on huggingchat, I'd even pay 30$ a month for it. In my opinion its perfect for translating. I would use it locally but I don't have a server that could run the full model and using quants will make the model worse.

I would like to pay 9$ per month for longer context + relaxed rate limit + unquantized usage of huggingface Chat

Hope you guys keep HuggingChat free forever ๐Ÿ™

As hugging face gives access to Host unlimited models, datasets, and Spaces for free.
Hope so Hugging Chat will remain free.

A famous Hindi Quote - "Umeed Pe duniya kayam Hai"

Translation - "The world is alive in hope."

Well, see what happens in future.

Upon closer inspection it seems like Nous-Hermes-2-Mixtral-8x7B-DPO is still a bit better than command r plus at translating from Chinese to english. It understands the meaning a bit more and especially writes it far better to read. I wonder how good the new 8x22 instruct model of mistral is gonna be. Anyway all the models are really good and have amazing uses! I hope we can access those that get released in the future too. Thank you very much for hosting them.

Hugging Chat org

Umeed Pe duniya kayam Hai

๐Ÿ’ฏ

@nsarrazin will assistant creators get a choice in which model to migrate to? i think this should be an option as recreation in another model is like starting anew.

a past comment:

  1. What will happen to the Assistant if a model is taken down? Migrate to new llm with context token +prompt as we/bot authors can change sys prompt of the assistants anytime? unlikely ig. or we could have a migration system for our old chats.
  2. shall there be a "View Sys Prompt" just like in regular chats beside/below the bot button? As the assistant button at the top shows the latest prompt only while the chat might have started with another prompt. (doesn't change the already active chat really)(once it recognized the changed sys prompt upon me mentioning only a part of it)

zephyr mixtral 8x 22b from hugging face comming soon ?
zephyr-orpo-141b-A35b-v0.1

What happened to the openchat model why was it removed

@SvCy 1. You can change to any llm even after making bot, or his llm was removed.

What happened to the openchat model why was it removed

Because very few people are using it.

Hugging Chat org

We just released HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1 on HuggingChat!

image.png

Try it out here: https://huggingface.co/chat/models/HuggingFace4/zephyr-orpo-141b-A35b-v0.1

Hugging Chat org

Shout out to @nicksuh who called it early ๐Ÿ˜…

What happened to the openchat model why was it removed

@Gerrytheskull models come and go.. nothing is permanent sadly.. besides OpenChat wasn't being used by that many of users i think. plus, new models were added.. command r+ and now zephyr

@SvCy 1. You can change to any llm even after making bot, or his llm was removed.

@KingNish oh users can change llms after creation now? sounds great.. thanks for the info!

@nsarrazin could you add model usage-over time graph on the model list page?
It would be more engaging and fun + new users can see what's trending.

@nsarrazin could you add model usage-over time graph on the model list page?
It would be more engaging and fun + new users can see what's trending.

  • feature like Assistant of the week (Like space have space of the week)

Is it possible to add the WizardLM-2-8x22B model to the available models?

image.png

Wizard is super competitor of current GPT4.

image.png

Wizard is super competitor of current GPT4.

Wizard seems like a killer model! We would love to see it on HuggingChat.

Wizard seems like a killer model! We would love to see it on HuggingChat.

There is only one big problem with this is that it has 141B parameters which makes it slow.

There is only one big problem with this is that it has 141B parameters which makes it slow.

The CohereForAI/c4ai-command-r-plus 110B params model works normally, so this should also work in normal mode. Additionally, there is the HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1 model with 141B params that also works quickly and is available in HuggingChat.

@CmetankaPY Ohh, i forget about them.

@CmetankaPY I found a discussion which stating that Zephyr has only 35b active parameters

https://huggingface.co/HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1/discussions/9

Did anyone notice that Zephyr 141B-A35B isn't even nearly as good as Command R+, despite having more parameters? I also noticed that some smaller models perform way better than Zephyr 141B-A35B.

Did anyone notice that Zephyr 141B-A35B isn't even nearly as good as Command R+, despite having more parameters? I also noticed that some smaller models perform way better than Zephyr 141B-A35B.

Because zephyr ha only 35b active parameters not 141b.
Read this for more info - https://huggingface.co/HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1/discussions/9

Please add AI generated images.

Please add AI generated images.

You can use image generation in chat using pollination

Some Example:-
https://hf.co/chat/assistant/6612cb237c1e770b75c5ebad
https://hf.co/chat/assistant/65bff23f5560c1a5c0c9dcbd

Hugging Chat org

๐Ÿšจ Meta Llama 3 70B is now available on Hugging Chat!

GLdkE2cXoAA5Y_X.jpeg

Let us know what you think about this one!

This comment has been hidden

Llama-3 seems great, but I expected it to beat GPT-4 ๐Ÿ˜…. So far can't see any open-source model that comes close to Command R+ performance

Llama-3 seems great, but I expected it to beat GPT-4 ๐Ÿ˜…. So far can't see any open-source model that comes close to Command R+ performance

Wizard Beated Command R+ and Even a very good competitor of ChatGPT

image.png

Llama-3 seems great, but I expected it to beat GPT-4 ๐Ÿ˜…. So far can't see any open-source model that comes close to Command R+ performance

Wizard Beated Command R+ and Even a very good competitor of ChatGPT

image.png

I believe Wizard will be the new open-source king, but I can't find it anywhere, I think Microsoft deleted it for some reason.

Hugging Chat org

I believe Wizard will be the new open-source king, but I can't find it anywhere, I think Microsoft deleted it for some reason.

what did Satya see

I believe Wizard will be the new open-source king, but I can't find it anywhere, I think Microsoft deleted it for some reason.

image.png

Hope to SEE Wizard on hugging face.

Hey victor could you adjust the repetition penalty for llama? Because Iโ€™m trying to do some creative writing but it literally gives me the same output every time I retry

Hey victor could you adjust the repetition penalty for llama? Because Iโ€™m trying to do some creative writing but it literally gives me the same output every time I retry

just do it yourself from advanced setting at the bottom of models name

Hey victor could you adjust the repetition penalty for llama? Because Iโ€™m trying to do some creative writing but it literally gives me the same output every time I retry

just do it yourself from advanced setting at the bottom of models name

IMG_0214.jpeg

This is all I see

This is all I see

Click create new assistant then you will able to see

image.png

The quality of Dolphin-Mistral/Mixtral of Cognitivecomputations is much better than that of Nous-Hermes, which may be a more suitable choice. I also used them in my own local ollama - until Command-R+ subverted the game.

P.S. Llama3 is so bad for my use. It is not even as good as the quantitative version of the above two models.

I just checked the model configuration of Command-R-Plus and noticed that the context window is limited. Is it because of cost consideration? If so, I hope to add a Q4 version for 128K-context -window supportโ€”โ€”and it should be much faster.

I just checked the model configuration of Command-R-Plus and noticed that the context window is limited. Is it because of cost consideration? If so, I hope to add the Q4 version for 128K-context -window supportโ€”โ€”and it should be much faster.

But what about quality, quantization decreases quality very much.

I just checked the model configuration of Command-R-Plus and noticed that the context window is limited. Is it because of cost consideration? If so, I hope to add the Q4 version for 128K-context -window supportโ€”โ€”and it should be much faster.

But what about quality, quantization decreases quality very much.

Then Q8? with extremely low Temp, Top_P and Top_K. In any case, the quality of command-R+ surpasses most models.

In addition, the impact of quantification on quality is not so devastating. The latest research can even be quantified with 1bit to achieve a nearly non-quantification effect.

Detailed review of Llama 3 70B:

Coding: 8/10

Capability: Llama 3 is capable of generating code snippets in various programming languages, including Python, Java, C++, and JavaScript. It can also help with code completion, debugging, and optimization.

Limitation: While it can generate code, it may not always be correct or efficient. It may also struggle with complex algorithms or nuanced programming concepts.

Example: I asked Llama3 to write 10 complex questions. It generated a correct solution for 9, but some of them were not the best one.

Creative Writing: 9/10

Capability: Llama 3 is capable of generating creative writing, including stories, poetry, and dialogues. It can understand context, tone, and style, and produce writing that is engaging and coherent.

Limitation: While it can generate creative writing, it may lack the nuance and depth of human-written work. It may also struggle with complex themes or abstract concepts.

Example: I gave 10 creative story generation tasks to him. It generated a engaging and well-structured story, but it lacked the emotional depth and complexity of a human-written work.

Multiple Language: 8.5/10

Capability: Llama 3 is capable of understanding and generating text in multiple languages, including English, Hindi, Chinses, Japanese, Spanish, French, German, Italian, and many others. It can also translate text from one language to another.

Limitation: While it can understand and generate text in multiple languages, it may not always be perfect in terms of grammar, syntax, or idiomatic expressions.

Example: I givee Llama 3 10 paragraphs of different languages to translate. It generated a accurate translation, but it lacked emotions, nuance and cultural context of a human.

General Knowledge: 9/10

Capability: Llama 3 has a vast knowledge base and can answer questions on a wide range of topics, including history, science, technology, literature, and more.

Limitation: While it has a vast knowledge base, it may not always be up-to-date or accurate. It may also struggle with abstract or nuanced concepts.

Example: I asked llama 3 about 10 diff complex GK questions . It generated a accurate and informative response, but it lacked the depth and nuance.

Maths: 6.5/10

Capability: llaama 3 is capable of solving mathematical problems, including algebra, geometry, calculus, and more. It can also help with mathematical concepts and theories.

Limitation: While it can solve mathematical problems, it may not always be able to explain the underlying concepts or find efficient approach and many times give wrong solutions.

Example: I asked Llama 3 to solve 10 complex high school problem. It generated a correct solution for 6 only, in 1 it follow right method at half and remaining 3 are purely incorrect.

Internet Search: 8/10

Capability: Llama3 can search the internet and provide relevant information on a wide range of topics. It can also help with finding specific information or answering complex questions.

Limitation: While it can search the internet, it may not always be able to evaluate the credibility or accuracy of the sources it finds.

Comparison with other models:

Llama 2
Llama 3 is a significant improvement over LLaMA 2 in terms of its capabilities and performance. It has a more advanced language model, better understanding of context and nuance, and improved generation capabilities. It is also more knowledgeable and accurate in its responses.
.
.
.
(More to be added)
.
.
.
Overall, Meta-Llama-3-70B-Instruct is a powerful and versatile language model that can perform a wide range of tasks and answer complex questions. While it has its limitations, it is a significant improvement over previous language models and has the potential to revolutionize the field of natural language processing.
.....................................................................................................
If you liked the review and want review for more models Give a thumbs up ๐Ÿ‘

Detailed review of Llama 3 70B:

Please do not use LLMs-style correct nonsense to describe the model's performance, thank you!

Note: Why do I think Dolphin performs better?

  • System prompt-free cross-language capabilities. When communicating in Chinese, Llama(1/2/3) or vanilla mistral 7B must be induced with system prompts to spit out fragmented Chinese. Nous-Hermes, CR+, and the Dolphin series do not have this problem.
  • Uncensored. Dolphin will never reject you.
  • It even has a programming-specialized version based on starcoder2.

Detailed review of Llama 3 70B:

Please do not use LLMs-style correct nonsense to describe the model's performance, thank you!

I wrote this entirely by myself, and you're claiming it's nonsense generated by LLM.

Repetition penalty for llama3 needs to be higher

I think we should add dolphin as itโ€™s a good model

noticed that current chats are not being named. can we assume it's under work for now?

Do you plan to release mistralai/Mixtral-8x22B-Instruct-v0.1 to the chat ? meta-llama/Meta-Llama-3-8B-Instruct could be also great.

Do you plan to release mistralai/Mixtral-8x22B-Instruct-v0.1 to the chat ? meta-llama/Meta-Llama-3-8B-Instruct could be also great.

Yeah, the instruct of 8x22 is AMAZING, Id like to use it over the chat too.

Do you plan to release mistralai/Mixtral-8x22B-Instruct-v0.1 to the chat ? meta-llama/Meta-Llama-3-8B-Instruct could be also great.

Yeah, the instruct of 8x22 is AMAZING, Id like to use it over the chat too.

Command-R-Plus is already overloading there. Is 8x22B really a reasonable choice? Llama3 8B can replace Mistral 7B and be the default configuration, anyway is broken now.
IMG_8372.jpeg

are all the models that come and go from huggingchat is open-sources?

Hugging Chat org

are all the models that come and go from huggingchat is open-sources?

yes sir

[New Model REQUEST] MTSAIR/MultiVerse_70B

This model outperforms Command R+, Llama 3 70B and many more, on open llm leaderboard.
As, command R+ is facing many issues. This model is a great alternative to command R+.
and It has only 70B parameters.
This model is currently #1 chat model on Open LLM leaderboard.

image.png

License - https://huggingface.co/MTSAIR/MultiVerse_70B/discussions/7#66278c8e430a12425331b183

Model Link - https://huggingface.co/MTSAIR/MultiVerse_70B

๐Ÿ‘ to support this model.
(Hugging Face team will add Model on Community Demand)

[New Model REQUEST] MTSAIR/MultiVerse_70B

It is based on Alibaba's Qwen72B, which means that it has been under severely censorship. Test scores sometimes don't make sense.

I suggest that Chinese models be treated with caution. They are never disappointing in terms of overfitting and Chinese political rights.

Conclusion: You'd better try this model before recommend it. Their Space is broken. On the other hand, quantifying or replacing Command-R+ with 35B Command-R is still a cost-effective choice.

For a full replacement, I would recommend this list of models:

  1. Command-R/Command-R+_Q6 or Q8
  2. Llama3 70B and subsequent versions with larger parameters
  3. Llama3 8B as a representative of small models and TASK_MODEL
  4. Phi-3-mini, can also be used as TASK_MODEL
  5. Dolphin/Nous-Hermes Mixtral 8x7B
  6. Anything else you want to add, such as Mistral-OpenOrca, Dolphin-Mistral, Qwen1.5... does not include vanilla Mistral or Mixtral 8x7B or Gemma, but Mixtral 8x22B is acceptable(better deploy with Q6).

*All the above quantitative suggestions are based on llama.cpp and gguf formats.

I suggest that Chinese models be treated with caution. They are never disappointing in terms of overfitting, just like their students.

@Mindires Hey, please treat every country and individual with respect. This is a community platform. So, Please do not spread hate or anything similar.

โ€œEverybody is a genius. But if you judge a fish by its ability to climb a tree, it will spend its whole life believing that it is stupid.โ€ โ€“ Albert Einstein

[New Model REQUEST] Microsoft/WizardLM-2

This model outperforms Command R+, Llama 3 70B, Mixtral 8x22B and many more.
And giving tough competition to - Claude 3, Gemini Ultra, GPT-4, etc.

image.png
image.png

License - Apache 2.0

Model Link - https://huggingface.co/alpindale/WizardLM-2-8x22B [Unofficial] (Official added soon)

๐Ÿ‘ to support this model.
(Hugging Face team will add Model on Community Demand)

[New Model REQUEST] Microsoft/WizardLM-2
-snip-

The legality of that is questionable, since Microsoft took it down.

[New Model REQUEST] Microsoft/WizardLM-2
-snip-

The legality of that is questionable, since Microsoft took it down.

It's not legally questionable. They released the model under the Apache 2.0 license, so anyone with a copy of the model can use, modify, and distribute it according to the license terms.

@EveryPizza Microsoft removed Wizard2 because it was uncensored.
So, they will post it again soon.

image.png

Microsoft removed Wizard2 because it was uncensored.

So they will censor it and release it again

Microsoft removed Wizard2 because it was uncensored.

So they will censor it and release it again

It's been a few days, and the censored version has now been released.

Review of Phi-3 Mini 4k Instruct:

Coding: 8.5/10

Capability: As Phi-3 is fine-tuned on High Quality Data of GPT-4. The performance is truly magical; According to his size of Just 3.8B. It excels in code completion, debugging, and optimization tasks, making it a valuable tool for developers.

Limitation: Phi-3 may occasionally produce code that is not optimal or entirely correct. It can encounter difficulties with complex algorithms or intricate programming concepts that require deep domain expertise.

Example: When tasked with creating 20 complex coding questions, Phi-3 delivered correct solutions for 19. However, some solutions were not the most efficient or elegant. But it Outperforms ChatGPT 3.5 (Free Version).

Creative Writing: 9/10

Capability: Phi-3 has a strong capability for creative writing, crafting stories, poetry, and dialogues with a clear understanding of context, tone, and style. Its outputs are engaging.

Limitation: Itโ€™s creative, but sometimes it doesnโ€™t hit the feels or the depth like something a person would write, especially with complex or deep themes.

Conclusion: Because of Dataset of GPT 4, It has vast advancement in creative writing.

Multiple Language Proficiency: 7/10

Capability: Phi-3 is capable of understanding and generating text in numerous languages, including English, Hindi, Chinese, Japanese, Spanish, French, German, Italian, and more.

Limitation: While Phi-3 is proficient in multiple languages, there are many lapses in grammar, syntax, or idiomatic expressions, which can detract from the authenticity of the text.

Example: Phi-3 translated 20 paragraphs from various languages with high accuracy. However, the translations manyimes missed the emotion and meaning of text.

General Knowledge: 9/10

Capability: Phi-3 has more knowledge as compare to its size. (It outperforms all 7b,13b and many 30b and some 70 b Models)

Limitation: Although its size is small. SO, Phi-3's information may not always be current or completely accurate. It can also struggle with detailed discussions on historical topics.

Example: Phi-3 was asked Different GK questions. It provided accurate and informative responses, but occasionally lacked the depth (Reason is his size).

Mathematics: 7/10

Capability: Phi-3 is proficient in solving mathematical problems, including those in algebra, geometry, calculus, and beyond. It can assist with understanding mathematical concepts and theories.

Limitation: Phi-3 may not consistently explain the underlying concepts clearly or choose the most efficient methods, and it can sometimes provide incorrect solutions.

Example: Phi-3 was tasked with solving 20 complex high school mathematics problems. It correctly solved 13, partially followed the right method for 3, but the remaining 4 were incorrect.

Internet Search: 8.5/10

Capability: Phi-3 can effectively search the internet to provide relevant information on a wide array of topics. It can assist in locating specific details or answering intricate questions.
....................................................................................................

Some useful Tips

  1. Phi3 + Internet > GPT 3.5
  2. Phi it is currently best model for local ai.
    ....................................................................................................

Comparison with other models:

Compared to Phi-2, Phi-3 represents a significant leap in handling complex tasks such as coding, mathematics, general knowledge, and creativity. It demonstrates an advancement in language model capabilities, offering a more sophisticated understanding of context and delivering highly knowledgeable and accurate responses.
(Compared to Phi 2)
....................................................................................................

Overall:

Phi-3 is a Magical model. We can see a wast difference between him and his competitors. It surpasses all 7b models and nearly all 13b models in performance. Eagerly waiting for the release of Phi-3 7B and 13B.

....................................................................................................

Thanks! to Microsoft for This high quality Model and hugging chat team to make it available free on HuggingChat

Fun Fact: HuggingChat team is very busy that they even forget to officially announce๐Ÿ˜… that Phi-3 is Available on HuggingChat.
So, Here is Link go Check it Out -> https://huggingface.co/chat/models/microsoft/Phi-3-mini-4k-instruct

......................................................................................................

If you find this review helpful and would like more reviews of similar models, please let me know! ๐Ÿ‘
You can follow me to get notified about next model Review.

See U in Next Review ๐Ÿค—

[New Model REQUEST] Microsoft/WizardLM-2

I created a Demo of WizardLM 2 7b model on Space,
Check it Out - https://huggingface.co/spaces/KingNish/WizardLM-2-7B

While many of the community members are requesting models based on the Open LLM scores. I believes that mods of this community also do have an eye on the open llm board. If a model seems a fit, they will surely add the model hopefully. We all want the best models to be present in the hugging face chat

I'm starting to face issues with Command R+; it's starting to hallucinate badly, doesn't follow requests properly, and gives one-word lazy answers even when I explicitly tell it to provide in-depth, expanded responses in the system prompt.

Here we can discuss about HuggingChat available models.

image.png

Is there a way to select another model other than the ones listed? Or, is there any other UI that someone could suggest me to deploy a model I fine-tuned myself previously? Thanks!

How can i add a new model by myself?

Hugging Chat org

How can i add a new model by myself?

By using chat-ui directly: https://github.com/huggingface/chat-ui

This comment has been hidden
Hugging Chat org

This is not the right place to post this @zoyahammad (here we discuss models on HuggingChat)

Sign up or log in to comment