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 monthsRecordings per month over the last 12 months — 1 in all, peaking in Aug 2026 with 1.
Appearances
Okay, so f so fine tuning.
So fightuning um um uh has multiple steps.
First is to um pick up an open source model and that means that all of these open source models are um documented, have uh um
technical documentation which I usually read to understand uh what is the architecture, what is the um how uh big or how large uh KV cache for example can um can be because it will impact on the size of the GPUs.
So I pick up a model, an open source model, let's say Quen, which is an excellent reasoning model, or Codlama, uh from Meta.
And then uh uh I prepare the dataset.
Data set basically is by far the most important part and the most tedious part.
Yes, is by far the most important component of the entire fine-tuning process system.
And the data set basically is um the first layer of analysis.
The f dataset is made of um let's say simply put uh question answers.
So the idea is that we we need to uh um teach the model um let's say what is OSPF using uh and what is an
area what is uh ABR etc.
using multiple uh question and answers which tackle various angles of the same um characteristic or parameter or uh network uh l aspect.
So the model should see the same questions in multiple ways in order to understand the pattern.
B basically this is how we learn um things by recognizing patterns.
And um this is the first phase, which is called uh supervised fine-tuning.
When this dataset is prepared and should have, uh, tens of thousands of entries.
Uh of course it should be also evaluated and tested.
There are multiple ways to see that.
Um
Showing 241–260 of 420 · page 13 of 21
← Previous
Next →