Article
PRX Part 3 — Training a Text-to-Image Model in 24h!
•
54
UygarUsta
RivianG
AI & ML interests
Computer Vision
Recent Activity
upvoted an article 5 days ago
PRX Part 3 — Training a Text-to-Image Model in 24h! liked
a model 14 days ago
cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit liked
a model 27 days ago
zai-org/GLM-OCR Organizations
None yet
upvoted an article 5 days ago
liked a model 14 days ago
liked a model 27 days ago
liked 2 models about 2 months ago
liked a model 3 months ago
reacted to hesamation's post with ❤️ 3 months ago
Post
3142
this is big... 50 AI researchers from Bytedance, Alibaba, Tencent, and other labs/universities just published a 300-page paper with surprising lessons about coding models and agents (data, pre and post-training, etc).
key highlights:
> small LLMs can beat proprietary giants
RL (RLVR specifically) gives small open-source models an edge over big models in reasoning. a 14B model trained with RLVR on high-quality verified problems can match the performance of OpenAI's o3.
> models have a hard time learning Python.
mixing language models during pre-training is good, but Python behaves different from statically typed languages. languages with similar syntax (Java and C#, or JavaScript and TypeScript) creates high positive synergy. mixing Python heavily into the training of statically typed languages can actually hurt because of Python's dynamic typing.
> not all languages are equal (coding scaling laws)
the amount of data required to specialize a model on a language drastically depends on the language. paper argues like C# and Java are easier to learn (less training data required). languages like Python and Javascript are actually more tricky to learn, ironically (you see AI most used for these languages :)
> MoE vs Dense (ability vs stability)
MoE models offer higher capacity, but are much more fragile during SFT than dense models. hyperparams in training have a more drastic effect in MoE models, while dense models are more stable. MoE models also require constant learning rate schedules to avoid routing instability.
> code models are "insecure" by default (duh)
training on public repos makes models learn years of accumulated insecure coding patterns. safety fine-tuning often fails to work much on code. a model might refuse to write a hate speech email but will happily generate a SQL-injection vulnerable function because it "works."
read the full paper:
From Code Foundation Models to Agents and Applications: A Practical Guide to Code Intelligence (2511.18538)
key highlights:
> small LLMs can beat proprietary giants
RL (RLVR specifically) gives small open-source models an edge over big models in reasoning. a 14B model trained with RLVR on high-quality verified problems can match the performance of OpenAI's o3.
> models have a hard time learning Python.
mixing language models during pre-training is good, but Python behaves different from statically typed languages. languages with similar syntax (Java and C#, or JavaScript and TypeScript) creates high positive synergy. mixing Python heavily into the training of statically typed languages can actually hurt because of Python's dynamic typing.
> not all languages are equal (coding scaling laws)
the amount of data required to specialize a model on a language drastically depends on the language. paper argues like C# and Java are easier to learn (less training data required). languages like Python and Javascript are actually more tricky to learn, ironically (you see AI most used for these languages :)
> MoE vs Dense (ability vs stability)
MoE models offer higher capacity, but are much more fragile during SFT than dense models. hyperparams in training have a more drastic effect in MoE models, while dense models are more stable. MoE models also require constant learning rate schedules to avoid routing instability.
> code models are "insecure" by default (duh)
training on public repos makes models learn years of accumulated insecure coding patterns. safety fine-tuning often fails to work much on code. a model might refuse to write a hate speech email but will happily generate a SQL-injection vulnerable function because it "works."
read the full paper:
From Code Foundation Models to Agents and Applications: A Practical Guide to Code Intelligence (2511.18538)
- 1 reply
liked a model 3 months ago
upvoted a collection 3 months ago
liked a Space 4 months ago
liked a model 4 months ago
updated a model 4 months ago
published a model 4 months ago
reacted to nouamanetazi's post with 🤗 4 months ago
Post
4543
After training 𝐒𝐦𝐨𝐥𝐋𝐌𝟑 on 𝟑𝟖𝟒 𝐇𝟏𝟎𝟎𝐬 for nearly a month, I've come to realize something most people overlook: 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐢𝐬 𝐭𝐡𝐞 𝐦𝐚𝐤𝐞-𝐨𝐫-𝐛𝐫𝐞𝐚𝐤 𝐟𝐚𝐜𝐭𝐨𝐫 𝐢𝐧 𝐋𝐋𝐌 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠. 🔥
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious 𝐍𝐂𝐂𝐋 𝐞𝐫𝐫𝐨𝐫𝐬, or when your expensive GPU cluster is running at 𝟔𝟎% 𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲, the problem isn't your model. It's most probably a 𝐦𝐢𝐬𝐮𝐬𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐡𝐚𝐫𝐝𝐰𝐚𝐫𝐞. 🛠️
Questions that seemed simple but had no clear answers: Why is 𝐌𝐨𝐄 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐬𝐥𝐨𝐰𝐞𝐫 𝐭𝐡𝐚𝐧 𝐝𝐞𝐧𝐬𝐞 𝐦𝐨𝐝𝐞𝐥𝐬? Which 𝐍𝐂𝐂𝐋 𝐟𝐥𝐚𝐠𝐬 should we actually set? How often should we checkpoint without killing throughput?
That's why we built 𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤 📖: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐥𝐚𝐲𝐞𝐫 that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: 𝐇𝐁𝐌𝟑 𝐡𝐢𝐭𝐭𝐢𝐧𝐠 𝟑 𝐓𝐁/𝐬, 𝐍𝐕𝐋𝐢𝐧𝐤 𝟒.𝟎 𝐫𝐞𝐚𝐜𝐡𝐢𝐧𝐠 𝟕𝟖𝟔 𝐆𝐁/𝐬, 𝐏𝐂𝐈𝐞 𝐆𝐞𝐧𝟒 𝐚𝐭 𝟏𝟒.𝟐 𝐆𝐁/𝐬. Then we ran collective operations across 𝟏𝟐𝟖 𝐆𝐏𝐔𝐬 (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from 𝟒𝟖𝟎 𝐆𝐁/𝐬 on a single node to 𝟑𝟐𝟎-𝟑𝟓𝟎 𝐆𝐁/𝐬 across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤: https://lnkd.in/e5MKXUHS
Shared with ❤️ by the HuggingFace team
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious 𝐍𝐂𝐂𝐋 𝐞𝐫𝐫𝐨𝐫𝐬, or when your expensive GPU cluster is running at 𝟔𝟎% 𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲, the problem isn't your model. It's most probably a 𝐦𝐢𝐬𝐮𝐬𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐡𝐚𝐫𝐝𝐰𝐚𝐫𝐞. 🛠️
Questions that seemed simple but had no clear answers: Why is 𝐌𝐨𝐄 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐬𝐥𝐨𝐰𝐞𝐫 𝐭𝐡𝐚𝐧 𝐝𝐞𝐧𝐬𝐞 𝐦𝐨𝐝𝐞𝐥𝐬? Which 𝐍𝐂𝐂𝐋 𝐟𝐥𝐚𝐠𝐬 should we actually set? How often should we checkpoint without killing throughput?
That's why we built 𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤 📖: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐥𝐚𝐲𝐞𝐫 that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: 𝐇𝐁𝐌𝟑 𝐡𝐢𝐭𝐭𝐢𝐧𝐠 𝟑 𝐓𝐁/𝐬, 𝐍𝐕𝐋𝐢𝐧𝐤 𝟒.𝟎 𝐫𝐞𝐚𝐜𝐡𝐢𝐧𝐠 𝟕𝟖𝟔 𝐆𝐁/𝐬, 𝐏𝐂𝐈𝐞 𝐆𝐞𝐧𝟒 𝐚𝐭 𝟏𝟒.𝟐 𝐆𝐁/𝐬. Then we ran collective operations across 𝟏𝟐𝟖 𝐆𝐏𝐔𝐬 (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from 𝟒𝟖𝟎 𝐆𝐁/𝐬 on a single node to 𝟑𝟐𝟎-𝟑𝟓𝟎 𝐆𝐁/𝐬 across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
𝐓𝐡𝐞 𝐒𝐦𝐨𝐥 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤: https://lnkd.in/e5MKXUHS
Shared with ❤️ by the HuggingFace team
liked 3 models 5 months ago
reacted to sergiopaniego's post with 🔥 5 months ago
Post
1497
Super nice intro to fine-tuning with TRL, just dropped by @google (runs free on Colab)!
They use SFT + QLoRA to fine-tune the tiny Gemma 3 270M model for emoji generation
Here’s what the fine-tuned model generates for the prompt: “I'm learning to tweet” → 🐦🗣💻
Colab: https://colab.research.google.com/github/google-gemini/gemma-cookbook/blob/main/Demos/Emoji-Gemma-on-Web/resources/Fine_tune_Gemma_3_270M_for_emoji_generation.ipynb
Try it out: google/emoji-gemma
Learn more: https://developers.googleblog.com/en/own-your-ai-fine-tune-gemma-3-270m-for-on-device/
They use SFT + QLoRA to fine-tune the tiny Gemma 3 270M model for emoji generation
Here’s what the fine-tuned model generates for the prompt: “I'm learning to tweet” → 🐦🗣💻
Colab: https://colab.research.google.com/github/google-gemini/gemma-cookbook/blob/main/Demos/Emoji-Gemma-on-Web/resources/Fine_tune_Gemma_3_270M_for_emoji_generation.ipynb
Try it out: google/emoji-gemma
Learn more: https://developers.googleblog.com/en/own-your-ai-fine-tune-gemma-3-270m-for-on-device/
reacted to hba123's post with 🔥 5 months ago
Post
4073
🤖 What if building your own robot arm costs less than £220?
For years, robotics has been locked behind high prices and complex systems.
So we decided to change that.
Today, we’re open-sourcing Ark-Bot — a fully 3D-printed, 6-DOF robot arm that works seamlessly with our Python robotics library, Ark.
And yes… It’s only £215.86 to build.
🧠ArkBot Specs 🧠
1️⃣ Reach: 1 meter
2️⃣ Weight: 2.6 kg
3️⃣ Payload: 1.8 kg 💪
4️⃣ DOF: 6
5️⃣ Input Voltage: DC 12V
🤟Fully 3D-printable & open-source
🤟Integrated with Ark — no ROS required
📹 We’ve also released a video showing the full assembly process — because robotics should be something everyone can learn, build, and improve on.
👩🎓 With Ark-Bot, anyone — from students to AI researchers — can experiment with embodied AI, robot learning, and control algorithms on real hardware, affordably.
If you could control a 1-meter robot arm from your laptop for under £220…
👉 What would you build first?
🔗https://github.com/Robotics-Ark/ark_bot
🎥 https://www.youtube.com/watch?v=Kuk4pC0EaEw&feature=youtu.be
For years, robotics has been locked behind high prices and complex systems.
So we decided to change that.
Today, we’re open-sourcing Ark-Bot — a fully 3D-printed, 6-DOF robot arm that works seamlessly with our Python robotics library, Ark.
And yes… It’s only £215.86 to build.
🧠ArkBot Specs 🧠
1️⃣ Reach: 1 meter
2️⃣ Weight: 2.6 kg
3️⃣ Payload: 1.8 kg 💪
4️⃣ DOF: 6
5️⃣ Input Voltage: DC 12V
🤟Fully 3D-printable & open-source
🤟Integrated with Ark — no ROS required
📹 We’ve also released a video showing the full assembly process — because robotics should be something everyone can learn, build, and improve on.
👩🎓 With Ark-Bot, anyone — from students to AI researchers — can experiment with embodied AI, robot learning, and control algorithms on real hardware, affordably.
If you could control a 1-meter robot arm from your laptop for under £220…
👉 What would you build first?
🔗https://github.com/Robotics-Ark/ark_bot
🎥 https://www.youtube.com/watch?v=Kuk4pC0EaEw&feature=youtu.be
- 2 replies
reacted to sergiopaniego's post with 🔥 5 months ago
Post
3010
A few days ago, Thinking Machines Lab released “LoRA Without Regret”, showing that LoRA can match full fine-tuning performance when configured right.
Naturally, we decided to reproduce the results with TRL and release a guide!
https://huggingface.co/docs/trl/main/en/lora_without_regret
Naturally, we decided to reproduce the results with TRL and release a guide!
https://huggingface.co/docs/trl/main/en/lora_without_regret