What is Prompt Tuning?
Вставка
- Опубліковано 21 гру 2024
- Explore watsonx → ibm.biz/BdvxRp
Prompt tuning is an efficient, low-cost way of adapting an AI foundation model to new downstream tasks without retraining the model and updating its weights. In this video, Martin Keen discusses three options for tailoring a pre-trained LLM for specialization, including: fine tuning, prompt engineering, and prompt tuning ... and contemplates a future career as a prompt engineer.
Get started for free on IBM Cloud → ibm.biz/sign-u...
Subscribe to see more videos like this in the future → ibm.biz/subscri...
#ai #watsonx #llm
Im stunned how Martin is able to write backwards on this board so efficiently
they just mirror the whole recording
haters will say that they just mirror the whole recording
seems you need to have that skill if you want to work at IBM
@@subusrable Almost right. What Martin is showing here is just the entry level. You actually have to be able to write in any direction. At IBM we call this 360-degree scribbling. And in any color, of course! (Patent pending - but we're considering to open-source the technology.)
At IBM Research we are even working on writing in n-dimensional space. Stay tuned.
Agility and flexibility are key!
Really like these summarization videos on this channel. While they do not go into depth, I appreciate the overarching concepts being outlined and put into context in a clean way without throwing overly specific stuff in the mix.
Which channels would you recommend that go into more depth?
Awesome content. Thanks for uploading.
It's great that the video calls out the differences between soft prompting and hard prompting. While soft prompts offer more opportunities for performance tuning, practitioners often face the following issues:
- Choosing between hard prompting with a more advanced, but closed, LLM versus soft prompting with an open-sourced LLM that is typically inferior in performance.
- Soft prompting is model dependent, and hard prompting is less so.
Excellent broad explanation of complex AI topics. One can then deep dive once a basic understanding is achieved ! Thank you
More important question, what type of smart/whiteboard are you using?? I love it!
See ibm.biz/write-backwards
how do you get the AI to generate that tunable soft prompt?
So how do I get to those "soft prompts"? Do you have to use prelabeled examples for that?
Soft Prompt example:
I want to make a certain image in Stable Diffusion, but i don't know the exact prompt i need to type to generate that image, so i ask ChatGPT to generate that prompt for me (describing the characteristics of that image to be generated).
ChatGPT outputs the prompt, in this case my Stable Diffusion soft prompt.
You should make a guide on FlowGPT / Poe that delves into operators, delimiters, markdown, formatting, and syntax. I've been experimenting on these sites for a while, and the things they can do with prompts are mind-blowing.
Can you give some examples? I have yet to be impressed, but im notably hard to impress.
This that soft prompt is basically a trainable parameters, which also undergoing backpropagation and its weights are updated? Just like LoRA method, where you attach new trainable parameters to the model and train only those new parameters.
Very concise and information, but tell me, what technology do you use to write backwards so fast? Do you flip the board in post-production?
Yes, see ibm.biz/write-backwards for details
Could you explain labeling done in fine tuning and prompt tuning?
Excellent job explaining key AI terms!
What data set for supervised learning is used in prompt tuning
Could you please outline the advantages and disadvantages of fine-tuning versus prompting in the context of large language models?
Hi, nice talk by the way, but what about some examples of soft turning, i understand is human unreadable, but how exactly you achieve that ? by writing some code ? extra tools ? plugins ? thanks a lot for your reply :)
These are learnable parameters added on top the base language models. This is called soft tuning one of the example is prefix tuning. These parameters are learned.
Excelente video, bom trabalho
Why can't we use a decoder to convert the soft prompt to text and this ways it's interpretable? I don't quite understand
How do you make these soft prompts ?
Could you please explain a little detail about the strings of numbers how those are indexed? Are those some sort of abstraction that we fully understand!
Very informative lecture is this one... Probably everyone should have a little expertise in prompt engineering skill in near future.
There are other embedding models that can take strings of concepts and transform them into embedding vectors (string of numbers). You can store those in a number of vector databases.
How do you discover the correct soft prompts?
I'm doing a project where I need to categorise the transaction details from transactional SMS to be output in JSON type. Can I prompt tuning or prompt engr with hard prompt?
What is unfancy design prompt?
Does anyone know how they do these videos where it appears that they are writing on the screen. That is so neat!
See ibm.biz/write-backwards
A lot to unpack here. Great job explaining.
I have one question about the difference between incontext learning and prompt tuning with hard prompts. Are they synonymous?
So is prompt engineering still a viable career (only because we’re in the infancy stages of widespread “commercial” use)…..of LLMs like ChatGPt.
crisp and informative
funny & informative 👏👏👏
Why isn't this more popular if it actually works? All I see is LORAs and RL methods.
Thanks for the explanation, but still how could someone succeed in prompt engineering practically
We are trying to create a chatbot using OpenAI API and the response should be limited to the specific topic and it should not respond to the user queries which are not related to the topic. What is the best way to achieve this ? Prompt engineering or prompt tuning ?
Fine tuning
Use a better LLM.
A joke by ChatGPT:
Why did the Large Language Model (LLM) turn down a job as a DJ?
Because it thought "Prompt Tuning" meant it would have to constantly change the music!
Why don’t I see this anywhere if it’s better than normal prompts
not fully understood except- prompt tuning-prompt engineering- hard tuning-soft tuning. :P
Hello, I have some splendid news that will bring a smile to your face!
I learned Python in 2 months, great language. Then, I learned the SQLs that Python plays well with. Then, it hit me: AI is doing most of this work! So what is there for me and you to do? "My career may be over before it's begun". Yes, indeed UNLESS we can start using Python for regular business processing, like Accounts Receivable/Payable, Inventory Management, Order Processing, etc. In other words, we can't all be doing AI, especially when it, itself, is doing AI, cheaper, faster, and better.
God no. Please grow up soon so you can comprehend that python is killing the internet.
Really like all of Martins videos but this one only explains what prompt-tuning is not.
Nomenclature is it's own problem.
As a newbee how come I be pro in propmt engineering
IBM could go for the plunge and make a Quantum computer with 10 million Quantum computer chips with 1000 Qubits and optical circuits instead of just one chip.
Chuning
Soft peompring is confusing
🙂
I agree. AI soft prompts are not readable
"A string of numbers is worth a thousand words" tf does that even mean?
Is he writing backwards
See ibm.biz/write-backwards
This is too abstract. Some concrete examples would have helped.
Wow, you managed to make an 8 minute video on prompt tuning without actually talking about what it is or how one would even begin to implement it. All I gleaned from this is that it has something to do with embeddings... Do better IBM...
I agree it’s a little obscure. I gave this a second watch through because your comment made me realize that I too wasn’t clear.
Here is what I’ve noted:
First step:
Model creation: A model is created by training it from tons of data (very expensive to do)
Because a model alone doesn’t work consistently at this point (racist, errors, hallucinations, toxic,…) it needs more work to be ready for the public. To make it ready one of the three strategies are used: fine tuning, prompt engineering, or prompt tuning with soft prompts. (All three could be used as well, I’ve read papers about such cases.)
Fine tuning :
Give you have a model, now you create examples about the domain the LLM will represent. The examples are labeled to help the model know what’s is going on. This strategy is labor intensive. (Labeling is another whole area to read up on.)
Prompt engineering:
Humans design prompts in a human language (explain to the model how to behave). Example: when I tell you a word in English, you respond with the word in French.
Prompt tuning using soft prompts:
Soft prompts are created by the AI using fine tuning data. These prompts are encoded (not human readable) into a vector.
The above is the first six minutes of the video. Next the lecturer show these three applications by adding them to the box picture. This is confusing because it seems like he is applying all three strategies but then concludes that prompt tuning gets the best results. So I guess he is saying use prompt tuning.
Since AiML is a new field, I think people will be applying many different strategies in order to get their models to work properly. And this is just scratching the surface. Every few months, people will come up with other strategies that improve the situation. 10 years from now a bunch of these strategies will be discarded and there will be other new ones. The field of ML is defining their design patterns. Pattern books will be written as solutions mature. Prompt engineering and prompt tuning are the two patterns he talks about.
I hope that helps. Thinking this through has certainly helped me so thanks for the prompt. 😊
yes. this is incredibly generic and communicates very little considering this is supposed to be from a communication theory expert.
Totally agrees..
I got one useful tidbit. That I have to stick the soft prompt into the embedding layer. How? Also unclear.
Shame not a single real-world example of prompt tuning isn't provided. I guess this video isn't about that kind of detail?
ChatGPT is not an LLM lol
yes it is