Hugging Face's logo
Activity Feed

AI & ML interests

None defined yet.

Recent Activity
codelionย  new activity 10 days ago
zero-gpu-explorers/README:ZeroGPU clone fails during tensor packing with Errno 28 โ€” 82.1 GB pack target
RioShiinaย  new activity 29 days ago
zero-gpu-explorers/README:All ZeroGPU Spaces stuck on "Starting" with clean logs โ€” CPU and paid GPU unaffected
codelionย  authored a paper about 2 months ago
An Empirical Study of Path Feasibility Queries
View all activity
sergiopaniegoย 
posted an update 2 days ago
view post
Post
3035
while preparing the last class of the Training Agents live series during the summer, i spent some time reading the post-training sections of many frontier model reports, to learn how they use RL environments to improve their models, and wrote a blog about it

if you use any kind of coding harness, or you saw the Blender scenes that went viral recently, this might be interesting to you

Blog: https://huggingface.co/blog/sergiopaniego/rl-environments-2026
appvoidย 
posted an update 3 days ago
view post
Post
88
We trained a 10.9M byte-level recurrent Transformer on L3 and L6. (Loop 3 and Loop 6)

Yet L4/L5 improved too, L8 held up, and the L3โ†’L6 gain grew during training.

Same weights. More compute. Better predictions.

This is a new architecture for effective compute after several steps beyond original training!

We mixed and matched components like time and mhc into an ouro-like byte-level language model and the result is BET, a byte-level step-elastic transformer that can run computation steps without significant degradation.

One of the coolest parts of this training was discovering how Gradient Descent decided to use the first layer as what we would consider a scratchpad! Totally destroyed for the decoder but somehow makes total sense for the next layer!

I believe looped-transformers are the future of edge computing and this is a first step towards it.

Blogpost: https://medium.com/@appvoidofficial/byte-level-elasticity-182fe2ed1d2f

appvoid/bet-10m
NCJย 
authored a paper 4 days ago
NCJย 
submitted a paper to Daily Papers 4 days ago
Aurelien-Morganย 
posted an update 7 days ago
view post
Post
2412
@retrain-pipelines execution engine is in perpetual evolution, with the aim to establish itself as SOTA, and for the long run.

However, we neglect no aspect of ML-Eng centricity.

If notebooks is where you like to do dev most,
we support you there 100% too.

Build crazy combos of inline tasks, deep parallel sub-DAG branches, nested asynchronous groups...

... the DAG renderer is undergoing an incremental upgrade

until the next one.

* starring toy tasks here. No ML has been hurt in this video ๐Ÿ™‚
appvoidย 
posted an update 8 days ago
view post
Post
3834
We got gpt6 before gta6
  • 15 replies
ยท
appvoidย 
posted an update 10 days ago
view post
Post
122
Any thoughts on Nvidia acquiring this website?

I don't know what to feel about it. But would be great if huggingface gets something similar to Kaggle with free GPU hours (or even days) for training.
  • 3 replies
ยท
codelionย 
in zero-gpu-explorers/README 10 days ago

ZeroGPU clone fails during tensor packing with Errno 28 โ€” 82.1 GB pack target

1
#185 opened 10 days ago by
Thananon
sergiopaniegoย 
posted an update 10 days ago
view post
Post
2035
Can you do RL over taste?

I've spent some time reproducing, in the open, Surya N's idea of training a model to paint with code. It's a coding model that learns to paint watercolours by writing JS code, trained with GRPO. I used TRL and OpenEnv for this, with the whole pipeline running on Hugging Face.

The interesting part is that the reward has no correct answer, unlike a math problem. In this case it's based on the artistic preferences of the person who builds the dataset.

Everything is published: the environment, the reference pool, the trained adapters, every painting of every run with the code that made it, and a write-up with all the decisions, including the ones that went wrong.

Blog post: https://huggingface.co/blog/train-to-paint-with-code
appvoidย 
posted an update 15 days ago
view post
Post
3299
Love how the small lm community is getting identity over time:

- Channel-Mixing
- XSA
- Three-tower
- Digit aware
- Loops

No one is doing the same! That's so cool.
  • 36 replies
ยท
sergiopaniegoย 
posted an update 16 days ago
view post
Post
451
catching up on some bookmarked reads from the summer, reading Antidoom from @liquidai

small reasoning models get stuck more easily when the task involves a long thinking trace and a hard problem. It starts repeating the same word over and over again ("Wait", "Alternatively"โ€ฆ), each repetition makes the next one likelier, and the generation is spent before it reaches an answer

they measured it, 10.2% of completions for an early LFM2.5-2.6B checkpoint and 22.9% for Qwen3.5-4B at greedy. After training those drop to 1.4% and 1.0%

the fix is FTPO (final token preference optimization). What I like is how narrow it is, it only touches the single token where the loop starts

three ways it differs from DPO:
> trains one token position, mid-generation, instead of whole sequences
> spreads probability across ~20 plausible alternatives instead of swapping one overtrained token for another
> keeps the regularizer in logit space, no softmax, so the rest of the vocabulary stays put

the third one is what makes it usable. If you want to edit one position without disturbing the model, you can't have a loss that reshuffles the other 150k logits on the way

and their explanation abt the result: the training teaches the model nothing new about math or code, it clears the failure mode that was blocking answers the model could already produce

full blog > https://www.liquid.ai/blog/antidoom

FTPO itself comes from Antislop, where it was built to strip overused phrasing. LiquidAI retargeted it to doom loops

and under the hood it's a subclass of TRL's DPOTrainer with compute_loss overridden, around 90 lines of loss and no new trainer

we documented that pattern in TRL's docs
https://huggingface.co/docs/trl/main/en/customization#change-the-training-objective
  • 1 reply
ยท
appvoidย 
posted an update 17 days ago
view post
Post
2642
Nobody knows what is doing, when you train a model, you are experimenting to advance the frontier, so keep failing ๐Ÿซต
  • 17 replies
ยท
appvoidย 
posted an update 19 days ago
view post
Post
2600
I hope that after this OpenAI disaster on Plus users, more people start realizing why Open Weights were always the only way.
  • 5 replies
ยท
appvoidย 
posted an update 22 days ago
view post
Post
1003
Byte-level state-space models. That sounded pretty scary for a scientist decades ago. Now we have:

1. Knowledge that deeper layers train smoothly.
2. Knowledge that Transformers work but is quadratic on sequence length.
3. Knowledge that SSMs work even better. Numerically unstable sometimes.
4. Speculative-decoding.
5. Open high-quality data.
6. Knowledge that KD works.

It slowly feels like is no longer a bad idea.
  • 3 replies
ยท
appvoidย 
posted an update 24 days ago
view post
Post
166
It's 2026 and there are no instantaneous/fast vision language models for cpus yet. That's another free idea.
  • 6 replies
ยท
sergiopaniegoย 
posted an update 24 days ago
view post
Post
310
super interesting new paper from Microsoft "Agent Lightning v1.0: Towards Harnessed Agentic RL" by Zhiyuan He et al.

same idea we've seen already several times: you train the agent inside the real harness it ships with, instead of a reimplementation of it

now that recipe has a name โ†’ harnessed agentic RL

paper: huggingface.co/papers/2608.17528

the tricky bit they nail down: one rollout is not one training sample

the harness calls the model many times, so a single episode โ†’ a variable number of (prompt, response) rows

you don't even know the batch size until the episode finishes running

its real contribution is being first to systematically map the four problems that fall out of that:

> retokenization + sample merging
> advantage calculation over a variable sample count
> loss normalization at the rollout level, not per sample
> backend scheduling when the batch size is dynamic

and it actually works โ†’ plain RL inside the real harness, no reimplementation

Qwen3.5-9B on SWE-bench Verified 41.8 โ†’ 56.4 (+14.6), with only ~6k examples

the whole thing is ~3,500 lines, any harness, self-hosted k8s

from our side, we've shared some materials on the same line you may want to check out :)

> Agentic RL: Token-In, Token-Out Done Right: https://huggingface.co/blog/huggingface/tito
> a full worked example, opencode owning its loop trained with GRPO: https://huggingface.co/blog/sergiopaniego/trl-openenv-harness-training
> Harness, Scaffold, and the AI Agent Terms Worth Getting Right: https://huggingface.co/blog/agent-glossary

on a similar line:

https://x.com/SergioPaniego/status/2062911580564496576
appvoidย 
posted an update 26 days ago
view post
Post
143
CaraArchive highlights a broader reality of putting data online: once something is publicly accessible, it becomes extremely difficult to guarantee that it will remain under your control.

If there is information or artwork that you absolutely do not want copied, archived, scraped, downloaded, or used by others, the safest option is still not to publish it publicly in the first place. That may sound obvious, but the internet was fundamentally designed to move and reproduce information, and there are countless ways to retrieve publicly accessible images:from ordinary browser tools and web scraping to automated or agentic systems.

That does not mean artists should simply accept every possible use of their work. Artists deserve meaningful control, attribution, compensation, and reasonable ways to express how their work may be used. But treating the technology and peopple using it itself as the enemy is unlikely to solve the underlying problem.

There probably isn't a technical solution that can make a publicly visible image simultaneously viewable by everyone and impossible to copy. The realistic goal should therefore be to create better norms, incentives, licensing systems, and tools around how that content is used.

Like, we can imagine a future where every artist gets his/her own credentials and some kind of fingerprint done just like blockchain works. But that requires substantial cooperation among organizations, companies and individuals.

Technology and art are not inherently opposing sides though.
RioShiinaย 
in zero-gpu-explorers/README 29 days ago

All ZeroGPU Spaces stuck on "Starting" with clean logs โ€” CPU and paid GPU unaffected

๐Ÿ‘ 2
3
#183 opened 30 days ago by
luna0805
appvoidย 
posted an update about 1 month ago
view post
Post
175
If you lack ideas for a cool model, here's one.

Train a model from scratch on wikipedia with one twist: the tokenizer changes the actual token ids used on every sample fed. If somehow still learns English, you have made an astonishing discovery.

You would have answered the question: Can a model learn human languages from structure alone?
  • 2 replies
ยท
appvoidย 
posted an update about 1 month ago
view post
Post
1971
Random corporate secret of tonight:

Try overfitting a tiny model on billions of high-quality datapoints: you can't. You can do 100 epochs and see the model still improving.

You're welcome.
  • 1 reply
ยท