Eduard (Edward) Dulharu

speaker

420 appearances 1 recordings 1 series first heard Aug 2026 last heard 7 Aug

Eduard (Edward) Dulharu’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months

Recordings per month over the last 12 months — 1 in all, peaking in Aug 2026 with 1.

Appearances

newest first · ▶ plays the moment

Heavy Networking · HN837: Agentic AI to Reduce MTTR · 7 Aug 2026 podcast

Which means that to be able to answer questions he never seen.

And I uh how how I do that is by the first thing which I'm doing, I'm preparing a test set, which I put it aside with questions the model will never see.

So I want to make sure that uh in the evaluation phase of the model.

Um he will go in is like an exam, yeah.

So the teachers they have preparing the uh I'm not gonna train you on the test.

Yeah, exactly.

Yeah.

You are training on

all the data available but you never see anything related to the test.

And this is the last step I'm or this is the uh is the first step I'm preparing the test.

I and then I uh prepare the data set which is a very tedious process.

It takes a lot of time even though I use also synthetic data, but uh I'm always in the look to check the accuracy because if

Uh the model sees only one uh entry uh data uh data point which has, I don't know, uh um wrong information about what is, I don't know, um ASBR in OSPF, it's enough then to hallucinate and to provide the wrong answer.

Right.

So and then you say it hallucinated and you don't know why.

Well, the reason is the data quality was uh

Not good enough.

So this is the first step, supervised fine-tuning.

Um so then I tested against this uh unseen uh test to see if it's an empirical uh validation basically.

I don't rely on losses, I don't rely on all the graphs which I see during the training, because during the training basically means to minimize the low loss function.

Showing 281–300 of 420 · page 15 of 21