Director, Learning Design & Development| PMIAsheville, NC, United States
Validating and checking outputs is critical when working with AI systems like Generative AI. Such validation approaches may include establishing clear criteria, implementing strong testing protocols, and continuous refinement.
In your experience with AI, what are some best practices for ensuring the results you receive are accurate, relevant, and aligned with your original goals?
1. Refine what you're asking--don't include unnecessary info
2. Be succinct
3. Know what it is you want the prompt to yield Saving Changes...
Elizabeth HalfordProject Manager| Cardea Project ManagementAlexandria, Va, United States
Jun 11, 2024 2:01 AM
Replying to Hakam Madi
...
This could be done by fine-tuning the chat context to fine-tuning the model using several strategies, such as Examples or few or many shots.
I'm currently working on a project. In my system Instruction [which Could be the scoping prompt if you are not accessing the API], I have the request and the verification method and criteria, so at the end of each output, I receive the confidence level achieved by AI.
With some training, I developed it further to output only results with an 85% confidence level or else provide an explanation or ask for clarification. This, btw, surprisingly jammed all the previous hallucinations.
Thanks for your post - very helpful! Saving Changes...
Hi! I believe that when you work with AI is essential you review and check if the results receive are accurate, like.
1. Define clear objectives;
2. Validate and Montor Outputs;
3. Use AI to Support, Not Replace, Human Judgment Saving Changes...
To get accurate, relevant, and goal-aligned results from AI, clearly define your objective using CREATE formula, provide context, fact-check critical details, iterate with feedback, specify the desired format and evaluation. This will help in getting accurate result. Saving Changes...
Valentine MrozekSenior IT Project Manager| Self Employed - Semi retiredMillersville, Md, United States
Easy...Randomly spot-check and verify.
and depending on the project / project size, etc...assign people from other parts of the project to spot-check (Verify) someone else's. In all cases everyone should a standard form of items across the board to check and an additional area for other items to verify. Date, print names and return to the PMO Office for final review. In some cases you may even assign to the vendors. (some have their own items to verify and you may be able to get copies of that).
Of course this can be done for all projects and this is in addition to what many people have stated about being clear, initial testing, etc,,,,
From my experience, AI tends to deliver better responses when I clearly define:
The role I expect it to play.
The goal I want to achieve.
The context surrounding the request.
The more detail I provide, the more precise and relevant the outcome.
Depending on the response I receive, I then clarify specific points, request refinements, or specify the format I want for the final answer.
I would usually stated the website/sources that I prefer the AI to refer to during prompting. Always recheck and validate, if not satisfied I will revise back my prompt as precise as possible, and give some example to refer. Saving Changes...
PMI Team Member
