Dominick Wirzba

Chronuid

1 4

dominick-wirzba-a46898115

AI & ML interests

None yet

Recent Activity

reacted to alexanderbering's post with 🔥 1 day ago

We just put our money where our manifesto is. Our position has been that the layer of an AI system that touches your data should be open and auditable, not something you rent on trust. So instead of asking you to believe that, we made it runnable. The ZenBrain Playground is a static HF Space that executes our open-source memory library, @zensation/algorithms (Apache-2.0, zero-dependency), live in your browser. No install, no backend, no mockup: it runs the actual published code. Four panels drive real functions from the library — Ebbinghaus-style retention curves, Hebbian strengthening/decay, sleep-consolidation replay + pruning, and Bayesian retrievability with confidence intervals. The library is vendored into the page, so what you run is the same code that's published on npm. ZenBrain is a neuroscience-inspired 7-layer memory architecture for AI agents. The open core is the 20-module algorithm library (FSRS, Hebbian, Ebbinghaus, sleep-consolidation, Bayesian confidence, and more). On LongMemEval-500 it reaches 91.3% of a long-context oracle's accuracy at 1/106 of the per-query token budget, and takes the highest mean rank across all 12 system-judge cells (4 systems x 3 LLM judges) against Letta, Mem0, and A-Mem. It ships with 11,589 automated tests. Full details are in the arXiv preprint. If you build agent memory, we'd genuinely like your eyes on it — open the Space, poke the algorithms, read the code, tell us where it breaks. Live demo: https://huggingface.co/spaces/zensation-ai/zenbrain-playground Model: https://huggingface.co/zensation-ai/zenbrain Paper: arXiv:2604.23878

reacted to albertvillanova's post with 🤗 2 days ago

🎉 KTO is now part of the stable TRL API As of Promote KTO to stable API, KTOTrainer and KTOConfig have graduated from trl.experimental to the stable trl API. https://github.com/huggingface/trl/pull/6175 This one closes out a long road. Over the past 6+ months, the "Align KTO with DPO" effort landed ~90 PRs methodically bringing KTO up to the standard we hold for stable trainers, one carefully-scoped change at a time: - Feature parity with DPO: full VLM support (incl. multi-image), sync_ref_model, PEFT + Liger, ZeRO-3 + PEFT dtype fix, pad_to_multiple_of, activation offloading, IterableDataset and dict eval_dataset, remove_unused_columns, and reference-logprob precomputation at init. - Consistency with DPO: aligned method order and signatures, tokenization, _prepare_dataset, PEFT handling, ref-model preparation for distributed training, and config layout — plus a new DataCollatorForKTO and output format. Metrics moved into _compute_loss and simplified to direct averages via the shared _metrics attribute. - Removing legacy baggage: dropped encoder-decoder support, BOS/EOS handling, null_ref_context, generate_during_eval, model_init, preprocess_logits_for_metrics, model/ref adapter names, and several dead config knobs. - Coverage: a full test suite mirroring DPO, text collator tests, VLM tests, and slow tests. - The promotion itself: the experimental → stable move (#6175) and shim cleanup (#6287), handled so downstream users get a clean deprecation path. Honestly, this has been one of the more complex tasks I've taken on since joining the team, not because any single change was hard, but because it demanded sustained consistency across a ~2,000-line trainer, with every branch, comment, and edge case kept in lockstep with DPO. Huge thanks to everyone who reviewed along the way (especially @qgallouedec), the incremental review cadence is exactly what kept this maintainable. KTO now sits on equal footing with our other flagship trainers. 🚀

reacted to albertvillanova's post with 🔥 2 days ago

View all activity

Organizations

reacted to alexanderbering's post with 🔥 1 day ago

Post

4105

We just put our money where our manifesto is.

Our position has been that the layer of an AI system that touches your data should be open and auditable, not something you rent on trust. So instead of asking you to believe that, we made it runnable.

The ZenBrain Playground is a static HF Space that executes our open-source memory library, @zensation /algorithms (Apache-2.0, zero-dependency), live in your browser. No install, no backend, no mockup: it runs the actual published code. Four panels drive real functions from the library — Ebbinghaus-style retention curves, Hebbian strengthening/decay, sleep-consolidation replay + pruning, and Bayesian retrievability with confidence intervals. The library is vendored into the page, so what you run is the same code that's published on npm.

ZenBrain is a neuroscience-inspired 7-layer memory architecture for AI agents. The open core is the 20-module algorithm library (FSRS, Hebbian, Ebbinghaus, sleep-consolidation, Bayesian confidence, and more). On LongMemEval-500 it reaches 91.3% of a long-context oracle's accuracy at 1/106 of the per-query token budget, and takes the highest mean rank across all 12 system-judge cells (4 systems x 3 LLM judges) against Letta, Mem0, and A-Mem. It ships with 11,589 automated tests. Full details are in the arXiv preprint.

If you build agent memory, we'd genuinely like your eyes on it — open the Space, poke the algorithms, read the code, tell us where it breaks.

Live demo: zensation-ai/zenbrain-playground
Model: zensation-ai/zenbrain
Paper: arXiv:2604.23878

reacted to albertvillanova's post with 🤗🔥 2 days ago

Post

3329

🎉 KTO is now part of the stable TRL API

As of Promote KTO to stable API, KTOTrainer and KTOConfig have graduated from trl.experimental to the stable trl API. https://github.com/huggingface/trl/pull/6175

This one closes out a long road. Over the past 6+ months, the "Align KTO with DPO" effort landed ~90 PRs methodically bringing KTO up to the standard we hold for stable trainers, one carefully-scoped change at a time:
- Feature parity with DPO: full VLM support (incl. multi-image), sync_ref_model, PEFT + Liger, ZeRO-3 + PEFT dtype fix, pad_to_multiple_of, activation offloading, IterableDataset and dict eval_dataset, remove_unused_columns, and reference-logprob precomputation at init.
- Consistency with DPO: aligned method order and signatures, tokenization, _prepare_dataset, PEFT handling, ref-model preparation for distributed training, and config layout — plus a new DataCollatorForKTO and output format. Metrics moved into _compute_loss and simplified to direct averages via the shared _metrics attribute.
- Removing legacy baggage: dropped encoder-decoder support, BOS/EOS handling, null_ref_context, generate_during_eval, model_init, preprocess_logits_for_metrics, model/ref adapter names, and several dead config knobs.
- Coverage: a full test suite mirroring DPO, text collator tests, VLM tests, and slow tests.
- The promotion itself: the experimental → stable move (#6175) and shim cleanup (#6287), handled so downstream users get a clean deprecation path.

Honestly, this has been one of the more complex tasks I've taken on since joining the team, not because any single change was hard, but because it demanded sustained consistency across a ~2,000-line trainer, with every branch, comment, and edge case kept in lockstep with DPO.

Huge thanks to everyone who reviewed along the way (especially @qgallouedec ), the incremental review cadence is exactly what kept this maintainable.

KTO now sits on equal footing with our other flagship trainers. 🚀

2 replies

reacted to pranavupadhyaya52's post with 🔥 2 days ago

Post

3532

After several weeks of experimenting, debugging, and iterating, I am excited to share WikiSmartBotLM.

WikiSmartBotLM is a compact decoder only language model built from the ground up as an educational and practical project. The goal was not simply to train another language model, but to create one that is easy to understand, modify, and experiment with while following many of the architectural ideas used in modern LLMs.

The model is built on a custom Transformer architecture featuring Rotary Positional Embeddings, RMSNorm, SwiGLU feed forward layers, grouped query attention, and an efficient autoregressive decoder optimized for local inference.

The repository includes the complete model implementation, configuration files, tokenizer integration, training pipeline, inference scripts, checkpoint conversion utilities, and examples that demonstrate how each component works together. Whether you want to understand the forward pass, train your own model, or build applications on top of WikiSmartBotLM, everything is designed to be approachable.
You can directly run the model via the models Huggingfaces space, which I've included in the post.

Model Repository: pranavupadhyaya52/Wiki-SmartBotLM-Instruct

I hope WikiSmartBotLM becomes a useful resource for anyone who enjoys learning by building. Feedback, issues, feature requests, and contributions are always welcome.
https://huggingface.co/spaces/pranavupadhyaya52/WikiSmartBot

1 reply

reacted to kasbsquall's post with 🔥 25 days ago

Post

4186

🔎 UX Crime Scene — every interface hides a crime.

Drop a screenshot of ANY website or app, and THE INSPECTOR — a film-noir detective — works it as a crime scene: he circles each UX flaw on the real pixels, names the charge, and files a verdict with a letter grade. A UX audit that plays like a detective thriller.

But the verdict is just the opening statement. Now it goes further:

⚖️ THE TRIAL — put the interface on trial. The guilty UI elements take the stand and defend themselves while the Inspector rules from the evidence.
🖼️ THE RECONSTRUCTION — one click and FLUX.2 Klein rebuilds the worst element FIXED, live. Before/after, on the real pixels.
🔊 THE VOICE — hear the verdict read aloud (Kokoro, local, no keys).
🚨 MOST WANTED — a public rogues' gallery. Book your case onto a shared board where the city's worst interfaces are ranked by their crimes. Booked by the public.

Three small models, all on Modal (scale-to-zero), none over 32B:
👁️ Qwen2.5-VL-7B (vision agent) · 🖼️ FLUX.2 Klein (reconstruction) · 🔊 Kokoro-82M (voice)

📊 Human-graded: 84% grounding / 92% valid charges.

▶️ Trailer: https://youtu.be/6u58YIEPrkA
📹 Full walkthrough: https://youtu.be/WyQbY0XJ_9E
🕵️ Try it: build-small-hackathon/ux-crime-scene

Built solo for #BuildSmallHackathon (Gradio × Hugging Face). Open the case — the Inspector is waiting.

1 reply

replied to eabdullin's post 27 days ago

Learning under pressure can be stressfull, I know. Going for a walk could help. Good luck with your PhD! 😄

reacted to mmhamdy's post with 🔥 28 days ago

Post

4981

It was supposed to be a failed experiment. Instead, it led to the discovery of one of the most intriguing phenomena in neural networks, simply because a researcher forgot to turn it off and left it running....for a week!

In 2022, researchers at OpenAI were studying how neural networks generalize from their training data. For this task, they were training small transformer models to perform modular arithmetic.

The thing is, neural networks are weird. When a model has an abundance of parameters (like neural nets), it can easily overfit. It essentially memorizes its training data, scoring a perfect 100% accuracy when tested on it, but remains completely clueless when faced with any new instances not present in the training set (close to 0 accuracy). It is like memorizing 1 + 2 = 3 without understanding the concept of addition, so if 2 + 3 wasn't in the training set, the model fails miserably!

Usually, when a model overfits like this, people just cut their losses, turn off the experiment, and move on with their lives.

But sometimes they forget. And that is exactly what happened to our researchers at OpenAI. A week later, they checked back in, and a miracle had happened!

They discovered Grokking (And no, this has nothing to do with xAI's Grok , the term was originally coined by sci-fi author Robert Heinlein to mean understanding something so deeply that it becomes part of you). Grokking is when a neural network suddenly and abruptly learns to generalize long after it has overfitted. Just take a look at the graph in the image below!

Spooky, right! I told you neural nets are weird!

4 replies

liked a dataset 30 days ago

wikimedia/structured-wikipedia

Viewer • Updated May 19 • 10.5M • 11.6k • 389

reacted to RiverRider's post with 🔥 about 1 month ago

Post

4847

SRT-introspect: Live Token-by-Token Readout of LLM Internal Reasoning

I have released SRT-introspect, a new public demonstration that makes the hidden reasoning process of a frozen large language model visible in real time.

The interface runs a Qwen-2.5-7B backbone equipped with the SRT Adapter and Activation Verbalizer. As the model generates each token, the system continuously measures divergence across attention heads, identifies high-signal moments, and translates the corresponding hidden-state object representations into natural-language verbalizations. You see exactly what the model is internally representing at the precise points where its computation is most active, complete with divergence scores, reflexivity estimates, and per-layer traces.

This is not a summary of the final output. It is a direct window into the model’s latent conceptual landscape, showing the dominant training-data attractors that activate even when the prompt asks for first-principles reasoning. The adaptive scheduler concentrates verbalizations precisely where the real internal work occurs, turning what used to be opaque black-box generation into observable, analyzable data.

The result is the clearest public demonstration yet that modern LLMs possess a rich, structured semiotic infrastructure that can now be audited without retraining or fine-tuning.

Try it:
RiverRider/srt-introspect

reacted to prithivMLmods's post with 🔥 about 2 months ago

Post

5602

I've made 8 Spaces in the Qwen-Image-Edit series, and out of them, 5 Spaces reached “Space of the Week”! A few Spaces are still topping the list even after many months.

Cumulatively, the series has crossed 8.2 million+ ZeroGPU runs and nearly 4 million visitors overall.

Thanks for all the community support! 🤗❤️

🔗 Spaces: https://huggingface.co/collections/prithivMLmods/image-generation-apps-collection

4 replies

reacted to danielhanchen's post with 👍 about 2 months ago

Post

2811

Qwen3.6 MTP is here! Run locally on 20GB RAM. ⚡️

MTP enables Qwen3.6 to generate ~1.4–2.2× faster with no accuracy change.

Qwen3.6-27B: unsloth/Qwen3.6-27B-MTP-GGUF
Qwen3.6-35B-A3B: unsloth/Qwen3.6-35B-A3B-MTP-GGUF
Guide: https://unsloth.ai/docs/models/qwen3.6#mtp-guide

2 replies

reacted to juiceb0xc0de's post with 🔥 about 2 months ago

Post

1510

Introducing the Gemma-4-E2B Brain Atlas, an interactive neural census of every layer, every head, 16 behavior categories in Google's flagship 2B model. We ran 184,320 probe prompts across 35 layers × 8 components and mapped what came back.

The Brain Atlas is an interactive tool that lets you explore the internal behavior of Google's Gemma-4-E2B model layer by layer, head by head. Pick a behavior category, pick a layer, and see exactly which components light up and which go quiet. The dataset is fully queryable if you want to go deeper.

The mapping combines multiple single-direction techniques run in parallel across every layer and component. Activation taxonomy (classifying each neuron by how broadly it fires across prompt categories), coactivation pair analysis (which neurons lock together and on what topics), F-stat behavioral separation (one-way ANOVA per feature across 16 behavior categories), per-head specificity scoring, and a full compliance probe pipeline using SVD, sparse decomposition, and variance analysis.

Here's what I found when I ran it.

The sharpest behavioral signal isn't at the output. It's Layer 0. Up projection hits F=22.7, nearly 2x anything in the final third of the network. The model does its behavioral sorting before it's barely started, then spends the next 34 layers… doing what exactly?

The gate has a lifecycle. 70% dormant at L1, highest in the model. Brutal sparsification at L23–26 (>58% silent). Then reopens. The final five layers are the most alive gates anywhere. The model's last act is a gate flare.
Layer 4 routes 5 projections to dim 448. One layer. One dimension. That's a topology highway.

Zero specialist neurons. Not one. 1.2M neurons analyzed. None fires exclusively on a single category. This model distributes everything.

🧠 Space: juiceb0xc0de/gemma-4-e2b-brain-atlas
📊 Dataset (1.3M rows, fully queryable): juiceb0xc0de/gemma-4-e2b-atlas

reacted to spillai's post with 🔥 about 2 months ago

Post

8769

mm-ctx – fast, multimodal context for agents.

LLM-based agents handle text incredibly well, but images, videos, or PDFs with visual content are hard to interpret. mm-ctx gives your CLI agent multi-modal skills.

Try it interactively in Spaces: vlm-run/mm-ctx

Readme: https://vlm-run.github.io/mm/
PyPI: https://pypi.org/project/mm-ctx
SKILL.md: https://github.com/vlm-run/skills/blob/main/skills/mm-cli-skill/SKILL.md

mm-ctx is meant to feel familiar: the UNIX tools we already love (find/cat/grep/wc), rebuilt for file types LLMs can't read natively and designed to work with agents via the CLI.
- mm grep "invoice #1234" ~/Downloads searches across PDFs and returns line-numbered matches
- mm cat <document>.pdf returns a metadata description of the file
- mm cat <photo>.jpg returns a caption of the photo
- mm cat <video>.mp4 returns a caption of the video

A few things we obsessed over:
⚡ Speed: Rust core for the hot paths
🏠 Local-first, BYO model: Uses any OpenAI-compatible endpoint: Ollama, vLLM/SGLang, LMStudio with any multimodal LLM (Gemma4, Qwen3.5, GLM-4.6V).
🔗 Composable: stdin + structured outputs
🤖 Drops into any agent via mm-cli-skills: Claude Code, Codex, Gemini CLI, OpenClaw.

We’d love to hear your feedback! Especially on the CLI and what file types and workflows you would like to see next.

2 replies

reacted to sergiopaniego's post with 🚀 2 months ago

Post

1903

OpenEnv is growing fast in tutorials. If you're looking to get started with RL environments, check them out

> evaluate your agents using OpenEnv
> learn how rewards work via rubrics
> connect agents via MCP
> many moreeeee!

anything you think it's missing?

https://meta-pytorch.org/OpenEnv/tutorials/index.html

reacted to danielhanchen's post with 🔥 2 months ago

Post

7780

We collaborated with NVIDIA to teach you how we made LLM training ~25% faster! 🚀

Learn how 3 optimizations help your home GPU train models faster:
1. Packed-sequence metadata caching
2. Double-buffered checkpoint reloads
3. Faster MoE routing

Guide: https://unsloth.ai/blog/nvidia-collab
GitHub: https://github.com/unslothai/unsloth

reacted to qgallouedec's post with 🔥 3 months ago

Post

2069

TRL v1.2 introduces the SSDTrainer 🚀

Simple Self-Distillation (SSD) from Apple's paper "Embarrassingly Simple Self-Distillation Improves Code Generation" is now available as an experimental trainer in TRL.

The recipe is as minimal as the name suggests: sample completions from the model itself at a training-time temperature, then fine-tune on those raw, unverified samples with plain cross-entropy. No reward model. No verifier. No teacher model. No reinforcement learning. Just prompts and the model.

from trl.experimental.ssd import SSDConfig, SSDTrainer

trainer = SSDTrainer(
    model="Qwen/Qwen3-4B-Instruct",
    args=SSDConfig(temperature=0.6, top_k=20, top_p=0.95),
    train_dataset=dataset,
)
trainer.train()

v1.2 also ships expanded tool-calling support (LLaMA 3.1 / 3.2, DeepSeek-V3), another round of KTO ↔ DPO alignment getting us closer to promoting KTO to stable, a big GRPO simplification for overlong tool results, deprecation of use_transformers_paged, and key fixes for VLM response parsing.

Full release notes: https://github.com/huggingface/trl/releases/tag/v1.2.0

reacted to kelsend's post with 👍 3 months ago

Post

3554

The rebuilt Hunyuan HY3 Preview is here!

I tested it on all the tricky scenarios where most LLMs usually face-plant—and guess what? It didn’t flop.

295B total params, 21B active params, 256K context window. Built on MoE architecture, it delivers trillion-parameter-level performance with a much smaller footprint. Long-context capabilities get a massive upgrade.

Agent abilities stand out this time: tool calling, workflow orchestration, and autonomous planning are far more stable in real business scenarios. AI PPT generation in Tencent Docs is also significantly smoother and more reliable.

Real-world tests on WorkBuddy show first-token latency down 54%, success rate over 99.99%, and an Agent workflow that ran continuously for 495 steps.

Its Coding Agent achieved top-tier results on both SWE-Bench Verified and Terminal-Bench 2.0

Now open-sourced on GitHub, HuggingFace, and ModelScope. Available on TokenHub at just 1.2 RMB per million tokens.

reacted to sergiopaniego's post with 🔥 3 months ago

Post

1443

Earlier this month, Apple introduced Simple Self-Distillation: a fine-tuning method that improves models on coding tasks just by sampling from the model and training on its own outputs with plain cross-entropy

And… it's already supported in TRL, built by Kashif Rasul. you can really feel the pace of development in the team 🐎

Paper by Ruixiang ZHANG, He Bai, Huangjie Zheng, Navdeep Jaitly, Ronan Collobert, Yizhe Zhang at Apple 🍎

How it works: the model generates completions at a training-time temperature (T_train) with top_k/top_p truncation, then fine-tunes on them with plain cross-entropy. no labels or verifier needed

You can try it right away with this ready-to-run example (Qwen3-4B on rStar-Coder):
https://github.com/huggingface/trl/blob/main/trl/experimental/ssd/ssd.py
or benchmark a checkpoint with the eval script:
https://github.com/huggingface/trl/blob/main/trl/experimental/ssd/ssd_eval.py

One neat insight from the paper: T_train and T_eval compose into an effective T_eff = T_train × T_eval, so a broad band of configs works well. even very noisy samples still help

Want to dig deeper?

Paper: Embarrassingly Simple Self-Distillation Improves Code Generation (2604.01193)
Trainer docs: https://huggingface.co/docs/trl/main/en/ssd_trainer

upvoted a paper 3 months ago

Embarrassingly Simple Self-Distillation Improves Code Generation

Paper • 2604.01193 • Published Apr 1 • 56

reacted to SeanLee97's post with 🔥 3 months ago

Post

8243

Our lab recently released a paper where we introduce ShadowPEFT, a new Parameter-Efficient Fine-Tuning (PEFT) paradigm tailored for edge computing scenarios.

Unlike traditional approaches such as LoRA and its variants, which inject trainable parameters directly into the weights of Transformer, requiring tight coupling with the backbone.

ShadowPEFT instead enhances the frozen large base model by adding a lightweight, centralized, pretrainable, and detachable Shadow network.
This shadow network operates in parallel with the base model, delivering learned corrections to each decoder layer. Because the shadow module is architecturally decoupled from the backbone, it can be independently trained, stored, and deployed, benefiting edge computing scenarios and edge-cloud collaboration computing.

- HF Paper: ShadowPEFT: Shadow Network for Parameter-Efficient Fine-Tuning (2604.19254)
- GitHub: https://github.com/ShadowLLM/shadow-peft
- HF Collection: https://huggingface.co/collections/shadow-llm/shadow-peft-models

7 replies

Dominick Wirzba

AI & ML interests

Recent Activity

Organizations

Chronuid's activity