No reviews yet. Be the first to share your experience!
Latest Posts
HAFSA
Mar 26, 2026, 11:16 PM
ስላዳመጣችሁኝ አመሠግናለሁ
Enough nerding for today will try to write properly about it if it ends up in a good state.
306
2
0
HAFSA
Mar 26, 2026, 11:16 PM
before the training searching up, I got something that said, 'If you see 1.3, then it goes down to 0.8, it means it's learning'
didn't even know what that meant but saw the numbers once it started training, and it showed 1.9, 1.8, 1.7
I was like yayyy is this actually running??
and I got an error (apparently, I was trying to load 60k+ pairs at once), so I adjusted stuff(too technical to write dekemegn)
and ended up fixing that.
172
2
0
HAFSA
Mar 26, 2026, 11:16 PM
I was scared when the number fluctuated between 1.7 to 1.6, then to 1.8, then again to 1.7 losses, but that's when I learned the term:
Stochasticity (randomness)
A lot of machine learning relies on controlled randomness. Things like sampling, initialization, and training batches introduce randomness, which actually helps the model generalize instead of memorizing everything.
so ig, it sees one message studies the pattern and finds it hard on some other messages so the loss increases..
again too technical for me
201
2
0
HAFSA
Mar 26, 2026, 11:16 PM
ended up finishing one batch of training. hope I resume it when possible and succeed.
201
2
0
HAFSA
Mar 26, 2026, 11:16 PM
📷 Photo
me naming my file 'chef' because I felt like I was cooking😭
154
7
HAFSA
Mar 26, 2026, 11:16 PM
cool things I discovered
Google Colab & Kaggle
I had heard about them before but never used them and they felt intimidating. Running code with a GPU there feels like a breath of fresh air compared to local setups. You can just stop worrying about your potato machine for a moment and focus on the process.
this was the main actor
Unsloth:
A toolkit that makes training large language models faster and lighter. It basically optimizes the training process so you can fine-tune models without needing ridiculous amounts of compute.
171
2
0
HAFSA
Mar 26, 2026, 11:16 PM
LoRA (Low-Rank Adaptation):
learnt this one after I saw it in a vibe code, and it said (this is where learning happens). I watched a video on it. again. very cool stuff, will prolly get back to it sometime later.
Instead of retraining a whole model, LoRA only trains small adapter layers. This means you can change how a model behaves (like its tone or style) while keeping the base model mostly untouched.
The learning process of the model
Training is surprisingly gradual. The model doesn’t suddenly “understand” things. It slowly adjusts its weights step by step, getting slightly better with every batch of data it sees.
157
2
0
HAFSA
Mar 26, 2026, 11:16 PM
an hour to research and export files, an hour to upload(shitty internet), an hour to understand what was being done, clean up data and implement everything, an hour more while the model trains(not even a huge dataset, idk what would have happened if it was the whole thing), and some frustrations with random errors, some hours watching tutorials on randomly unrelated topics, and running out of free-tiers of everything(you have to be rich to run rich tech that makes you richer), and some youtube
ended up spending 8 hours without moving from my position. the day has ended, but so has my sanity and health
136
1
0
HAFSA
Mar 26, 2026, 11:16 PM
you have to be rich to run rich tech that makes you riche
143
1
0
HAFSA
Mar 26, 2026, 11:16 PM
I never post things I'm unsure about because I have a debilitating fear of failure, but things I learnt today were so cool that it's worth it even if I fail
137
2
0
HAFSA
Mar 23, 2026, 11:15 PM
ignoring the narcissistic reasons I wanted something that was me, sounds, talks, and behaves like me
then come ai, apparently we can train models to be ourselves, and I've never dealt with anything ai related before unless from a user's perspective, so I wanted to try it. 'it's EASY' I was told
so I spent my day learning and experimenting on fine-tuning a model, which is only done halfway but still a really fun learning process
122
3
0
HAFSA
Mar 16, 2026, 01:24 PM
How ro reduce screentime if you use your phone to watch or listen to yt or things like, but not too much of scrolling
95
1
0
HAFSA
Mar 16, 2026, 01:24 PM
to spam or to die
92
1
0
HAFSA
Mar 16, 2026, 01:24 PM
a long long day
haven't coded or done anything related to it in a while
as someone famously said 'it's hard to care'
but found a new area of interest to play with
89
2
0
HAFSA
Mar 16, 2026, 01:24 PM
https://gurshaplus.com/hafsa
might have to join the hustle😔
106
0
0
HAFSA
Mar 16, 2026, 01:24 PM
Introducing https://www.gurshaplus.com/, an Ethiopian version of Buy Me a Coffee, built by our very own https://t.me/Natnael_Adane
It’s a platform where creators can receive support from their community.
Aaand If you enjoy the things I build or share, you can support me and give me a Gursha using the link below 💙
https://gurshaplus.com/mrnaty
34
0
0
HAFSA
Mar 16, 2026, 01:24 PM
https://gurshaplus.com/hafsa might have to join the hustle😔
109
2
0
HAFSA
Mar 16, 2026, 01:24 PM
learning is fun once you remove the reward part and just count the process as the reward
103
4
0
HAFSA
Mar 14, 2026, 10:24 PM
How ro reduce screentime if you use your phone to watch or listen to yt or things like, but not too much of scrolling
95
0
0
HAFSA
Mar 14, 2026, 10:24 PM
"You can not revel in the simplicity unless you remember the alternatives"
160
1
0
HAFSA
Mar 14, 2026, 10:24 PM
📷 Photo
https://bookpulse.lozi.me/ is a social reading platform where people can read books together
You can create or join book clubs, follow a shared reading schedule, and discuss ideas with others as you move through the pages.
If you prefer reading quietly, that’s also possible. You can create a private club with just your friends or even read completely on your own.
96
5
HAFSA
Mar 14, 2026, 10:24 PM
📷 Photo
160
5
0
HAFSA
Mar 14, 2026, 10:24 PM
Sudden cheerfulness after two days of 8 hr sleep
healthy?
125
1
0
HAFSA
Mar 14, 2026, 10:24 PM
📷 Photo
🚀 Introducing http://Dataset.ET/
The future of AI should speak Ethiopian languages.
Today we are launching http://Dataset.ET/ — an open community initiative to build the largest dataset for Ethiopian languages.
Why this matters:
Most AI systems today barely understand Amharic, Afaan Oromo, Tigrinya, Somali, and many other Ethiopian languages. Without datasets, AI will ignore our languages.
http://Dataset.ET/ is changing that.
What we are building:
• Speech datasets
• Text datasets
• Community-validated language corpora
• Open infrastructure for Ethiopian AI
How you can help:
1️⃣ Record sentences using our Telegram bot
2️⃣ Validate recordings from other contributors
3️⃣ Help expand datasets for Ethiopian languages
No technical skills needed — just your voice and your language.
🤖 Start contributing:
https://t.me/dataset_et_bot
🌍 Learn more:
https://dataset.et/
💬 Join the community:
https://t.me/dataset_et
Together we can teach AI to understand Ethiopia.
📢 Community leaders & channel admins:
If you run a Telegram group or community and believe Ethiopian languages should be part of the future of AI, we would really appreciate it if you could share this post with your community.
Your support can help thousands of people contribute their language and voice to this initiative. Thank you!
HAFSA
Mar 14, 2026, 10:24 PM
am I an open-source contributor now
110
0
0
HAFSA
Mar 14, 2026, 10:24 PM
I'm spamming on Nova's channel today
(ignore the hypocrisy of not even posting on my own channel)
193
5
0
HAFSA
Mar 14, 2026, 10:24 PM
📷 Photo
117
1
0
HAFSA
Mar 14, 2026, 10:24 PM
!!
Anyone who knows Addis based companies working on these fields:
electronics, nanotechnology, networking, communications, robotics, AI, communications and related fields
