Sare

cat lover 🐱

[Sare @xinkxlack]

too much oiia cat 💚💚💚💚 #OIIAOIIA

[Sare @xinkxlack]

#EastEnders

[XertXetin @xertxetin]

CetinLM is not a fine-tune of an existing model.

We’re building the entire training factory around it from scratch: the data pipeline, tokenizer, packing, training runtime, checkpointing, validation, generation diagnostics, and the systems used to measure whether learning is actually happening.

That also makes the project harder to explain sometimes.

People naturally compare it to the workflows they already know — fine-tuning an existing model, swapping a dataset, changing a seed, running a benchmark.

Our setup is different.

We own the full training path and the full failure surface.

If something breaks, there is no upstream model vendor to blame and no hidden training stack to fall back on. We have to trace the behavior through the whole system, from data to gradients to validation to generation.

That is why we don’t optimize around one scary metric or celebrate one good-looking number.

We evaluate the factory as a whole.
CetinLM is not just a model we are training.

We are building the machinery that creates the model.

#CetinLM #LLM #AI #MachineLearning

[Sare @xinkxlack]

#WWE2K27

[XertXetin @xertxetin]

CetinLM is not a fine-tune of an existing model.

We’re building the entire training factory around it from scratch: the data pipeline, tokenizer, packing, training runtime, checkpointing, validation, generation diagnostics, and the systems used to measure whether learning is actually happening.

That also makes the project harder to explain sometimes.

People naturally compare it to the workflows they already know — fine-tuning an existing model, swapping a dataset, changing a seed, running a benchmark.

Our setup is different.

We own the full training path and the full failure surface.

If something breaks, there is no upstream model vendor to blame and no hidden training stack to fall back on. We have to trace the behavior through the whole system, from data to gradients to validation to generation.

That is why we don’t optimize around one scary metric or celebrate one good-looking number.

We evaluate the factory as a whole.
CetinLM is not just a model we are training.

We are building the machinery that creates the model.

#CetinLM #LLM #AI #MachineLearning

[Sare @xinkxlack]

#PINTOSAWARDSㅤㅤ #ometv #nolimitt #monkeyapp𝗠𝗘𝗦𝗦𝗔𝗚𝗘

[AMBERK @GreenAIdev]

Yerli yapay zeka modeli CetinLM tam gaz gelişiyor! 🚀 1B akıllı base model çalışması büyük iddialar peşinde değil; çevre dostu, sürdürülebilir ve mini tüketimle en iyi sonucu hedefleyen, A'dan Z'ye sıfırdan yapılan harika bir Ar-Ge projesi. Takibe değer! 🌿🤖 @xertxetin #AI

[Sare @xinkxlack]

@xertxetin @oiiaoiiameme

[Sare @xinkxlack]

#kpsslisans 👏👏👍

[AMBERK @GreenAIdev]

Yerli yapay zeka modeli CetinLM tam gaz gelişiyor! 🚀 1B akıllı base model çalışması büyük iddialar peşinde değil; çevre dostu, sürdürülebilir ve mini tüketimle en iyi sonucu hedefleyen, A'dan Z'ye sıfırdan yapılan harika bir Ar-Ge projesi. Takibe değer! 🌿🤖 @xertxetin #AI

[AMBERK @GreenAIdev]

Yerli yapay zeka modeli CetinLM tam gaz gelişiyor! 🚀 1B akıllı base model çalışması büy��k iddialar peşinde değil; çevre dostu, sürdürülebilir ve mini tüketimle en iyi sonucu hedefleyen, A'dan Z'ye sıfırdan yapılan harika bir Ar-Ge projesi. Takibe değer! 🌿🤖 @xertxetin #AI

[XertXetin @xertxetin]

@neerajjj6785 🤣 This is how we are right now. #CetinLM 🫨 @grok

[Sare @xinkxlack]

#AEWAllIn

[XertXetin @xertxetin]

A clarification about CetinLM:
We are not optimizing for “the best tiny model.”
We are building a model-development methodology from zero.

CetinLM-1B is not our product.
It is our laboratory.

We are not trying to build a cute 1B model that runs on a consumer GPU.

We are building, from first principles, the machinery required to create an entirely new generation of language models.

Architecture.
Tokenizer.
Data.
Training systems.
Kernel profiling.
Scientific diagnostics.
Context scaling.
Post-training.
Reasoning.
Inference.

And we are documenting the experiment while it is happening — including the failures.

Most people see “1B parameters.”

We see a controlled research scale:
small enough to inspect deeply,
large enough for real language-model behavior to emerge.

At 573M tokens, Paris moved from rank 211 → 2.
Ankara moved from rank 6 → 1.
Validation is still breaking new lows.

We profiled attention backends instead of trusting defaults.
Audited a billion-token corpus for EOS boundaries.
Rejected optimizations that looked clever but lost empirically.
Kept the ones that survived measurement.

This is the part of AI people usually see only after billions of dollars, giant clusters and polished model releases have hidden the machinery underneath.

We are building that machinery in public.

The thesis is bigger than CetinLM-1B:

How much intelligence can disciplined architecture, data and systems engineering extract from every parameter and every watt of compute?

And if we can understand that at 1B...

what happens when the same methodology reaches 7B, 70B, or something much larger?

Large labs have enormous compute budgets. We do not.

That makes the experiment more interesting, not less.

Our thesis is that before scale becomes the answer, there is a huge amount of capability left on the table in architecture, data quality, training systems and post-training.

If that thesis survives at 1B, we scale the methodology — not just the parameter count.

CetinLM-1B is the test bench.
CetinLM is the research program.

#AI #LLM #MachineLearning #DeepLearning

[Sare @xinkxlack]

#BINI_SIGNALS_BROOKLYN

[XertXetin @xertxetin]

CetinLM-1B continues training. 🚀

We’ve now passed 20.5K steps / 167.9M tokens in the current base-model run.

📉 Validation loss: 4.115
📊 Perplexity: 61.26
🧠 Parameters: 1.048B
🎯 Target: 1B training tokens
🖥️ Development target: RTX 4070 Ti SUPER 16GB

We also started checkpoint-based benchmarking so we can measure the model as it learns — not just wait until the end.

Right now it is still a raw foundation model, so repetition and strange generations are expected. Chat, instruction following, code, reasoning, tools and web are later stages.

From here, every major checkpoint will be measured across:

Language → Knowledge → Turkish → Math → Logic → Code → Instruction → Conversation → Reasoning → Tools

The goal remains the same:

Build capable AI that doesn’t require enormous hardware.

This isn’t about making the biggest model possible.

It’s about seeing how much capability we can build, measure and improve within hardware people can realistically access.

Still training. Still learning. 🧠

#AI #LLM #MachineLearning #DeepLearning

[Sare @xinkxlack]

#BBBattleRoyale 👏🙏

[XertXetin @xertxetin]

We’re building CetinLM from the ground up.

Not a wrapper. Not a fine-tuned open model. Not a demo.

We are designing our own AI stack end-to-end:

• Custom multilingual tokenizer — 65,536 vocab
• Custom ~1B parameter Transformer architecture
• GQA + RoPE + SwiGLU + RMSNorm
• Custom data preparation & exact token-quota pipeline
• 1B training-token multilingual corpus
• Validation, checkpointing & resumable training
• BF16 + 8-bit optimizer + memory-efficient training for 16GB GPUs

Our current milestone: CetinLM-1B Base is actively training.
The goal is not to make a model with a large parameter count just for the number.
We want to build a foundation we can continue to develop:

Base → Instruction → Conversation → Code → Reasoning → Tool Use → Web

Eventually, CetinLM should be able to understand context, solve problems, write code, reason through tasks and use external tools when necessary.
We’re building the infrastructure first, then we’ll keep pushing the model forward.

This is only the beginning.

#CetinLM #AI #MachineLearning #LLM #ArtificialIntelligence #BuildInPublic

[Sare @xinkxlack]

#BBBattleRoyale 🤝👍

[XertXetin @xertxetin]

CetinLM-1B continues training. 🚀

We’ve now passed 20.5K steps / 167.9M tokens in the current base-model run.

📉 Validation loss: 4.115
📊 Perplexity: 61.26
🧠 Parameters: 1.048B
🎯 Target: 1B training tokens
🖥️ Development target: RTX 4070 Ti SUPER 16GB

We also started checkpoint-based benchmarking so we can measure the model as it learns — not just wait until the end.

Right now it is still a raw foundation model, so repetition and strange generations are expected. Chat, instruction following, code, reasoning, tools and web are later stages.

From here, every major checkpoint will be measured across:

Language → Knowledge → Turkish → Math → Logic → Code → Instruction → Conversation → Reasoning → Tools

The goal remains the same:

Build capable AI that doesn’t require enormous hardware.

This isn’t about making the biggest model possible.

It’s about seeing how much capability we can build, measure and improve within hardware people can realistically access.

Still training. Still learning. 🧠

#AI #LLM #MachineLearning #DeepLearning

[Sare @xinkxlack]

@xertxetin 🤣🥰

[Sare @xinkxlack]

@xertxetin cetinlm-1b-base current state;

[Sare @xinkxlack]

@xertxetin Tony Stark was able to build this in a cave! With a box of scraps!

[XertXetin @xertxetin]

We’re building CetinLM from the ground up.

Not a wrapper. Not a fine-tuned open model. Not a demo.

We are designing our own AI stack end-to-end:

• Custom multilingual tokenizer — 65,536 vocab
• Custom ~1B parameter Transformer architecture
• GQA + RoPE + SwiGLU + RMSNorm
• Custom data preparation & exact token-quota pipeline
• 1B training-token multilingual corpus
• Validation, checkpointing & resumable training
• BF16 + 8-bit optimizer + memory-efficient training for 16GB GPUs

Our current milestone: CetinLM-1B Base is actively training.
The goal is not to make a model with a large parameter count just for the number.
We want to build a foundation we can continue to develop:

Base → Instruction → Conversation → Code → Reasoning → Tool Use → Web

Eventually, CetinLM should be able to understand context, solve problems, write code, reason through tasks and use external tools when necessary.
We’re building the infrastructure first, then we’ll keep pushing the model forward.

This is only the beginning.

#CetinLM #AI #MachineLearning #LLM #ArtificialIntelligence #BuildInPublic

[XertXetin @xertxetin]

Got bored at 3 AM and decided to build an LLM from scratch.
Wrote the tokenizer, data pipeline, model and training system myself.

It’s training now. 👽

No idea how far this thing will go.
But we’re about to find out. 🧠🔥

#AI #LLM #MachineLearning #CetinLLM

[Sare @xinkxlack]

@xertxetin 🥰🤣

[XertXetin @xertxetin]

Wow, I randomly came across the test page for my own songs while just browsing around. 😂
https://t.co/rBigVF5z1p
It plays one of my songs and gives you a few different titles to choose from. Let’s see if you can actually guess the right one. 😂
Honestly, the test is pretty fun. Give it a try!
#OIIAOIIA #quiz #MusicLegend

[Sare @xinkxlack]

@oiiaoiiameme

[Sare @xinkxlack]

@xertxetin 😎