burtenshaw (ben burtenshaw)

posted an update 8 days ago

Post

1657

Qwen 3 Fine tuning >> MoE. Update the experiment thread to include config and script for fine-tuning the Qwen3-30B-A3B model.

The goal is to make a low latency non-thinking model for a daily driver coding, so 3 billion parameters active should be perfect.

✔️ training running
✔️ evals running
⏭️ improve dataset

The moe isn't going to fit into colab's A100 even with quantization (🙏 @UnslothAI ). So I've been working on HF spaces' H100s for this. Everything is available in the tread and I'll share more tomorrow.

burtenshaw/Qwen3-Code-Lite#1

posted an update 15 days ago

Post

2340

The rebooted LLM course starts today with an overhauled chapter 1 on Transformers:

👉 Follow the org to join the course:

huggingface-course

We’re starting from the foundations of modern generative AI by looking at transformers. This chapter is expanded in depth and features so contains new material like:

FREE and CERTIFIED exam on fundamentals of transformers
deeper exploration of transformer architectures and attention mechanisms
end -to-end exploration of inference strategies for prefill and decode steps

The course has leveled up in complexity and depth, so this a great time to join in if you want to build you own AI models.

posted an update 22 days ago

Post

1886

Hacked my presentation building with inference providers, Cohere command a, and sheer simplicity. Use this script if you’re burning too much time on presentations:

🔗 https://github.com/burtenshaw/course_generator/blob/main/scripts/create_presentation.py

This is what it does:
- uses command a to generates slides and speaker notes based on some material.
- it renders the material in remark open format and imports all images, tables, etc
- you can then review the slides as markdown and iterate
- export to either pdf or pptx using backslide

🚀 Next steps are: add text to speech for the audio and generate a video. This should make Hugging Face educational content scale to a billion AI Learners.

1 reply

·

reacted to yjernite's post with 🔥 22 days ago

Post

3183

Today in Privacy & AI Tooling - introducing a nifty new tool to examine where data goes in open-source apps on 🤗

HF Spaces have tons (100Ks!) of cool demos leveraging or examining AI systems - and because most of them are OSS we can see exactly how they handle user data 📚🔍

That requires actually reading the code though, which isn't always easy or quick! Good news: code LMs have gotten pretty good at automatic review, so we can offload some of the work - here I'm using Qwen/Qwen2.5-Coder-32B-Instruct to generate reports and it works pretty OK 🙌

The app works in three stages:
1. Download all code files
2. Use the Code LM to generate a detailed report pointing to code where data is transferred/(AI-)processed (screen 1)
3. Summarize the app's main functionality and data journeys (screen 2)
4. Build a Privacy TLDR with those inputs

It comes with a bunch of pre-reviewed apps/Spaces, great to see how many process data locally or through (private) HF endpoints 🤗

Note that this is a POC, lots of exciting work to do to make it more robust, so:
- try it: yjernite/space-privacy
- reach out to collab: yjernite/space-privacy

reacted to merve's post with 🔥 25 days ago

Post

4438

sooo many open AI releases past week, let's summarize! 🤗
merve/april-11-releases-67fcd78be33d241c0977b9d2

multimodal
> Moonshot AI released Kimi VL Thinking, first working open-source multimodal reasoning model and Kimi VL Instruct, both 16B MoEs with 3B active params (OS)
> InternVL3 released based on Qwen2.5VL, 7 ckpts with various sizes (1B to 78B)

LLMs
> NVIDIA released Llama-3_1-Nemotron-Ultra-253B-v1 an LLM built on Llama 405B for reasoning, chat and tool use
> Agentica released DeepCoder-14B-Preview, fine-tuned version of DeepSeek-R1-Distilled-Qwen-14B on problem-test pairs, along with the compiled dataset
> Zyphra/ZR1-1.5B is a new small reasoning LLM built on R1-Distill-1.5B (OS)
> Skywork-OR1-32B-Preview is a new reasoning model by Skywork

Image Generation
> HiDream releases three new models, HiDream I1 Dev, I1 Full, and I1 fast for image generation (OS)

*OS ones have Apache 2.0 or MIT licenses

4 replies

·

posted an update about 1 month ago

Post

3157

NEW UNIT in the Hugging Face Reasoning course. We dive deep into the algorithm behind DeepSeek R1 with an advanced and hands-on guide to interpreting GRPO.

🔗

reasoning-course

This unit is super useful if you’re tuning models with reinforcement learning. It will help with:

- interpreting loss and reward progression during training runs
- selecting effective parameters for training
- reviewing and defining effective reward functions

This unit also works up smoothly toward the existing practical exercises form @mlabonne and Unsloth.

📣 Shout out to @ShirinYamani who wrote the unit. Follow for more great content.

1 reply

·

replied to clem's post about 2 months ago

Open source extensions for Roombas would be so good! Like, could I add an arm and a trailer to my Roomba so it can pickup large objects whilst vacuuming.

reacted to clem's post with ❤️ about 2 months ago

Post

3746

Should we assemble affordable open-source robots at Hugging Face for the community. Would you buy them? At what price?

8 replies

·

posted an update about 2 months ago

Post

3823

The Hugging Face Agents Course now includes three major agent frameworks!

🔗

agents-course

This includes LlamaIndex, LangChain, and our very own smolagents. We've worked to integrate the three frameworks in distinctive ways so that learners can reflect on when and where to use each.

This also means that you can follow the course if you're already familiar with one of these frameworks, and soak up some of the fundamental knowledge in earlier units.

Hopefully, this makes the agents course as open to as many people as possible.

3 replies

·

posted an update about 2 months ago

Post

2511

The open LLM leaderboard is completed, retired, dead, ‘ascended to a higher plane’. And in its shadow we have an amazing range of leaderboards built and maintained by the community.

In this post, I just want to list some of those great leaderboards that you should bookmark for staying up to date:

- Chatbot Arena LLM Leaderboard is the first port of call for checking out the best model. It’s not the fastest because humans will need to use the models to get scores, but it’s worth the wait. lmarena-ai/chatbot-arena-leaderboard

- OpenVLM Leaderboard is great for getting scores on vision language models opencompass/open_vlm_leaderboard

- Ai2 are doing a great job on RewardBench and I hope they keep it up because reward models are the unsexy workhorse of the field. allenai/reward-bench

- The GAIA leaderboard is great for evaluating agent applications. gaia-benchmark/leaderboard

🤩 This seems like such a sustainable way of building for the long term, where rather than leaning on a single company to evaluate all LLMs, we share the load.

3 replies

·

posted an update about 2 months ago

Post

2230

Still speed running Gemma 3 to think. Today I focused on setting up gpu poor hardware to run GRPO.

This is a plain TRL and PEFT notebook which works on mac silicone or colab T4. This uses the 1b variant of Gemma 3 and a reasoning version of GSM8K dataset.

🧑‍🍳 There’s more still in the oven like releasing models, an Unsloth version, and deeper tutorials, but hopefully this should bootstrap your projects.

Here’s a link to the 1b notebook: https://colab.research.google.com/drive/1mwCy5GQb9xJFSuwt2L_We3eKkVbx2qSt?usp=sharing

1 reply

·

reacted to nicolay-r's post with 🔥 about 2 months ago

Post

1586

📢 With the recent release of Gemma-3, If you interested to play with textual chain-of-though, the notebook below is a wrapper over the the model (native transformers inference API) for passing the predefined schema of promps in batching mode.
https://github.com/nicolay-r/nlp-thirdgate/blob/master/tutorials/llm_gemma_3.ipynb

Limitation: schema supports texts only (for now), while gemma-3 is a text+image to text.

Model: google/gemma-3-1b-it
Provider: https://github.com/nicolay-r/nlp-thirdgate/blob/master/llm/transformers_gemma3.py

1 reply

·

replied to their post about 2 months ago

move this over to article, so I can keep it up to date: https://huggingface.co/blog/burtenshaw/gemma3-thinking

posted an update about 2 months ago

Post

2027

everybody and their dog is fine-tuning Gemma 3 today, so I thought I'd do a longer post on the tips and sharp edges I find. let's go!

1. has to be install everything form main and nightly. this is what I'm working with to get unsloth and TRL running

git+https://github.com/huggingface/transformers@main
git+https://github.com/huggingface/trl.git@main
bitsandbytes
peft

plus this with --no-deps

git+https://github.com/unslothai/unsloth-zoo.git@nightly
git+https://github.com/unslothai/unsloth.git@nightly

2. will brown's code to turn GSM8k into a reasoning dataset is a nice toy experiment https://gist.github.com/willccbb/4676755236bb08cab5f4e54a0475d6fb

3. with a learning rate of 5e-6 rewards and loss stayed flat for the first 100 or so steps.

4. so far none of my runs have undermined the outputs after 1 epoch. therefore, I'm mainly experimenting with bigger LoRA adapters.

from trl import GRPOConfig

training_args = GRPOConfig(
    learning_rate = 5e-6,
    adam_beta1 = 0.9,
    adam_beta2 = 0.99,
    weight_decay = 0.1,
    warmup_ratio = 0.1,
    lr_scheduler_type = "cosine",
    optim = "adamw_8bit",
    logging_steps = 1,
    per_device_train_batch_size = 2,
    gradient_accumulation_steps = 1,
    num_generations = 2,
    max_prompt_length = 256,
    max_completion_length = 1024 - 256,
    num_train_epochs = 1,
    max_steps = 250,
    save_steps = 250,
    max_grad_norm = 0.1,
    report_to = "none",
)

5. vision fine-tuning isn't available in TRL's GRPOTrainer, so stick to text datasets. but no need to load the model differently in transformers or Unsloth

from transformers import AutoModelForImageTextToText

model = AutoModelForImageTextToText.from_pretrained("google/gemma-3-4b-it)

if you want an introduction to GRPO, check out the reasoning course, it walks you through the algorithm, theory, and implementation in a smooth way.

reasoning-course

2 replies

·

posted an update about 2 months ago

Post

2112

Here’s a notebook to make Gemma reason with GRPO & TRL. I made this whilst prepping the next unit of the reasoning course:

In this notebooks I combine together google’s model with some community tooling

- First, I load the model from the Hugging Face hub with transformers’s latest release for Gemma 3
- I use PEFT and bitsandbytes to get it running on Colab
- Then, I took Will Browns processing and reward functions to make reasoning chains from GSM8k
- Finally, I used TRL’s GRPOTrainer to train the model

Next step is to bring Unsloth AI in, then ship it in the reasoning course. Links to notebook below.

https://colab.research.google.com/drive/1Vkl69ytCS3bvOtV9_stRETMthlQXR4wX?usp=sharing

4 replies

·

posted an update 2 months ago

Post

3864

I’m super excited to work with @mlabonne to build the first practical example in the reasoning course.

🔗

reasoning-course

Here's a quick walk through of the first drop of material that works toward the use case:

- a fundamental introduction to reinforcement learning. Answering questions like, ‘what is a reward?’ and ‘how do we create an environment for a language model?’

- Then it focuses on Deepseek R1 by walking through the paper and highlighting key aspects. This is an old school way to learn ML topics, but it always works.

- Next, it takes to you Transformers Reinforcement Learning and demonstrates potential reward functions you could use. This is cool because it uses Marimo notebooks to visualise the reward.

- Finally, Maxime walks us through a real training notebook that uses GRPO to reduce generation length. I’m really into this because it works and Maxime took the time to validate it share assets and logging from his own runs for you to compare with.

Maxime’s work and notebooks have been a major part of the open source community over the last few years. I, like everyone, have learnt so much from them.

replied to their post 2 months ago

thanks. back up now.

replied to their post 2 months ago

Thanks. Should be fixed now.

posted an update 2 months ago

Post

5653

I made a real time voice agent with FastRTC, smolagents, and hugging face inference providers. Check it out in this space:

🔗 burtenshaw/coworking_agent

9 replies

·

posted an update 2 months ago

Post

6438

Now the Hugging Face agent course is getting real! With frameworks like smolagents, LlamaIndex, and LangChain.

🔗 Follow the org for updates

agents-course

This week we are releasing the first framework unit in the course and it’s on smolagents. This is what the unit covers:

- why should you use smolagents vs another library?
- how to build agents that use code
- build multiagents systems
- use vision language models for browser use

The team has been working flat out on this for a few weeks. Led by @sergiopaniego and supported by smolagents author @m-ric .

ben burtenshaw

AI & ML interests

Recent Activity

Organizations

burtenshaw's activity