Building on HF

10 1 44

Dean Byrne PRO

Quazim0t0

dean-byrne-02a28b191

AI & ML interests

Joined HuggingFace: 1-20-2025 / DaisyChainAI🌼 / SmallLM's / San Francisco

Recent Activity

new activity about 13 hours ago

Intel/low_bit_open_llm_leaderboard:My models failed

updated a Space about 17 hours ago

Quazim0t0/neural-64-v2

published a Space about 17 hours ago

Quazim0t0/neural-64-v2

View all activity

Organizations

reacted to goumsss's post with 👀 about 20 hours ago

Post

641

🦁 I built a math app for my daughters — and they already love it.

NumZoo is a mental math trainer for kids. Answer questions, earn cute AI-generated animal rewards. My girls picked their favorite animals, chose a magical place, and haven't stopped playing since.

▶️ Watch the demo:
https://www.loom.com/share/c565cba49f2c4ad1bfb428e38ff4b629
🎮 Try it:
build-small-hackathon/numzoo

Stack:
- 🎨 FLUX.2-klein-4B — 4B params, Apache 2.0, 4 inference steps
- 🖼️ Custom kawaii style LoRA (trained on FLUX.2-klein-base-4B)
- ⚡ ZeroGPU · Gradio · Diffusers

The LoRA: locking in a kawaii storybook style

Out of the box, FLUX.2-klein is unpredictable — same prompt, wildly different styles. Not great when you want every reward to feel cozy and on-brand for a 6-year-old.

So I fine-tuned a style LoRA: 54 hand-curated scenes (12 animals × 10 places), captioned with a NUMZOO. trigger and no style words — the model learns the look from the images alone. Rank 32, 1500 steps, trained on a Modal A100 in ~45 min.

Published here: goumsss/numzoo-flux2-klein-lora

Two performance tricks that made it feel instant:
- Pre-generation: image generation starts in the background the moment the quiz begins. By the time a reward is earned (3 correct answers), the image is already ready. Zero wait.
- Cold-start fix: from_pretrained runs at module scope on CPU, outside the ZeroGPU budget. Inside @GPU we only do .to("cuda") — shaves the cold-start from potential timeout to ~1s.

Built for the 🤗 Build Small Hackathon. Small model, real joy. 🐾

reacted to TerminalCalm's post with 👍 about 20 hours ago

Post

805

A simple demo of TemperCheck, which is meant to quickly scan a screengrab of a social media profile and give the user a rough back-of-the-envelope estimation of how much of a temper they have. Made for the build small hackathon, relying on a Gemma4 edge model.

Made for the build-small-hackathon with ZeroGPU, Gradio, and Gemma.

Try it: build-small-hackathon/TemperCheck

1 reply

reacted to nevmenandr's post with 🔥 about 20 hours ago

Post

887

🔥 New Russian Stylometry Dataset!

Russian Stylometric Dataset (RSD) — 322 texts from the 19th – early 20th centuries (16 million words), prepared for analysis in stylo (R) and machine learning (Python).

📚 What's inside?

Fiction, journalism, scientific texts, drama, poetry

Grouped by author, gender, age, genre, literary movements (Romanticism/Realism)

Character speech (Tolstoy, Gogol, Ostrovsky)

Generated texts (LSTM, GPT)

📊 Use cases: authorship attribution, clustering, classification, benchmarking methods.

🔓 Public domain + GPL-3.0 license.

👉 Learn more: https://github.com/nevmenandr/RSD

DOI: 10.5281/zenodo.20701309

reacted to ykirpichev's post with 🔥 about 20 hours ago

Post

849

Glass-Box Agent for Build Small Hackathon.

A tiny ReAct-style agent where the trace is the interface: click a thought, retry a branch, label weak/useful nodes, and export preference pairs for DPO/RL-style training.

Space: build-small-hackathon/glass-box-agent
Demo: included in the Space at assets/glass-box-agent-demo.mp4
Track: An Adventure in Thousand Token Wood

#BuildSmallHackathon #Gradio #SmallModels

reacted to kriyanshi's post with 👍 about 20 hours ago

Post

904

Turn everyday phone requests into structured Android automation skills.

Flow: your prompt → fine-tuned Qwen2.5-3B classifier → recorded UI trajectory
15 skills: alarms, Spotify, WhatsApp, Slack, Uber, Gmail, YouTube, contacts, and more
Stack: Unsloth QLoRA on Modal · Modal inference API · Gradio Space

Live demo: build-small-hackathon/android-skill-router
Video: https://youtube.com/shorts/IQRHf7HfTDA

Track: Backyard AI · Sponsor: Modal · Achievements: Tiny Titan, Agent

1 reply

reacted to loay's post with 👍 about 20 hours ago

Post

982

I built EchoYard for the

build-small-hackathon : a tiny listen-and-repeat language practice app.

Pick a language, level, and voice style, listen to a short reference voice, record yourself, then get simple speaking feedback and a next practice step.

Built with

openbmb VoxCPM2 for multilingual reference audio and MiniCPM5-1B for friendly feedback.

Try it here: https://huggingface.co/proxy/build-small-hackathon-echoyard.hf.space

Would love feedback, especially on the recording flow and how useful the speaking tips feel.

reacted to Kasualdad's post with 👍 about 20 hours ago

Post

1028

From Plain English to DuckDB SQL: Building LFEDS
🏫 I just shipped Local First Education Data Stack— a plain-English-to-SQL assistant for school district analytics — for the HF Build Small Hackathon.

The problem: school staff have useful data (attendance, grades, enrollment, discipline) but no fast, private way to ask questions. Most AI tools send that data to a cloud API. LFED doesn't.

What it does:
→ Type a question like "What's the average GPA for chronically absent students in 2023-2024?"
→ A fine-tuned Qwen2.5-Coder-14B model generates DuckDB SQL
→ A validation layer rejects anything that isn't a SELECT
→ Results come back as a summary, table, CSV download, and the SQL itself

Two flavors:
- Live Space demo: transformers + PEFT on HF ZeroGPU
- Local-first: llama.cpp + GGUF Q4_K_M on your own machine — no data leaves

The fine-tune:
- 27,859 synthetic NL→SQL pairs
- Unsloth QLoRA r=32 on Qwen2.5-Coder-14B
- Trained on Modal A10G

Hardest lessons were not model training:
1. Scope the model's job tightly — schema + few-shots + SELECT only.
2. Validate before executing. Always.
3. ZeroGPU is PyTorch-only; llama.cpp won't work there.
4. Gradio's scoped Svelte CSS beats generic selectors — inspect the live DOM.
5. modal deploy + fn.spawn() is fire-and-forget; modal run dies if your terminal drops.
6. Data artifacts matter as much as the model — Parquet seeds, dataset card, model card.

I also published the training dataset: 25,886 question→SQL pairs on the Hub.

Links:
Demo: https://youtu.be/cE0yp4qmFIA
- Live Space: build-small-hackathon/Kasualdad_LFED
- LoRA adapter: build-small-hackathon/lfed-qwen2.5-coder-14b-sql-lora
- GGUF: build-small-hackathon/lfed-qwen2.5-coder-14b-sql-gguf
- Dataset: build-small-hackathon/lfed-training-data

#BuildSmallHackathon #BackyardAI #HuggingFace #TextToSQL #DuckDB #LocalFirst #EdTech #Qwen #QLoRA #LLM

reacted to ovi054's post with ❤️ about 20 hours ago

Post

3361

Qwen3-14B Manim Expert LoRA

For "Build Small Hackathon", I built a Gradio app that turns any concept into a Manim explainer video.

This is powered by Qwen3-14B + Manim LoRA I trained on a synthetic 10k dataset I generated.

👉 Try it now: build-small-hackathon/anim-vid-ai

2 replies

reacted to visqet's post with 👍 about 20 hours ago

Post

1072

Submitted to the build-small-hackathon! https://huggingface.co/proxy/build-small-hackathon-docintel.hf.space/

reacted to Moneyparking's post with 👍 about 20 hours ago

Post

1048

AI Bracket War Room 2026 — Build Small Hackathon demo

A fan-made World Cup 2026 matchday command center built through an iterative Codex-assisted workflow.

Official Space:
build-small-hackathon/ai-bracket-war-room-2026

Demo:
https://youtu.be/35rWi2RTueU

GitHub:
https://github.com/moneyparking/wc2026_planner

Unofficial fan-made project. No official affiliation claims. No real-money contest logic.

reacted to mmhamdy's post with 👍 about 20 hours ago

Post

1176

What if you could train a model on just 10 images instead of 60,000 and still get close to the same performance?

Traditional machine learning requires thousands, even millions, of data points to achieve high accuracy. But what if we could "distill" the entire dataset into just a few synthetic samples?

This is what Dataset Distillation offers. Unlike traditional knowledge distillation, we keep the model fixed and distill the knowledge contained in a massive training set into a tiny set of synthetic distilled images.

The goal is to train a model on this ultra-small set and achieve performance that almost matches what the same model would get when trained on the massive original dataset.

For example, training on only 10 distilled MNIST images (this is equivalent to a single image per class) yields 94% accuracy, compared to 99% when training on the full 60,000 images.

Interestingly, these distilled images look significantly different (as you can see in the image below) from natural images because they are optimized for model training rather than for matching the correct data distribution.

But that's not all.

Most importantly, this same method opens the door to a potent form of data poisoning. Because distilled images are specifically optimized for rapid learning, an attacker can create a tiny set of adversarial distilled images to cause a well-trained model to forget or misclassify a specific category.

What I find fascinating about dataset distillation is this: it mimics human-like learning by letting a model grasp a concept from a single example, but it does so using alien synthetic images that mean absolutely nothing to a human eye!

What about you? What are your thoughts on it?

2 replies

reacted to kanaria007's post with 👍 about 20 hours ago

Post

122

✅ Article highlight: *Institutional Memory & Forgetting for Learning Worlds* (art-60-172, v0.1)

TL;DR:
This article argues that if a living world becomes training data, memory becomes infrastructure.

Logs, dialogue, labels, releases, feature stores, and model weights can turn a world into something that cannot honestly forget. 172 makes deletion, redaction, exclusion, forgetting requests, SANITIZED/PUBLIC releases, and unlearning claims into receipted governance lifecycles.

Read:
kanaria007/agi-structural-intelligence-protocols

Why it matters:
• prevents learning worlds from becoming “unforgettable worlds”
• separates deletion, redaction, and future extraction exclusion
• makes right-to-be-forgotten requests caseable and appealable
• preserves canon facts without preserving every memory surface
• blocks public promises like “guaranteed deletion everywhere”

What’s inside:
• retention policy contracts for what may be kept, copied, trained on, or released
• corpus segment manifests and propagation indexes for known controlled copies
• forgetting request, adjudication, remedy, deletion, redaction, and exclusion receipts
• tombstone manifests and semantic preservation receipts for canon-safe forgetting
• use eligibility receipts for deciding whether a segment may train a future run
• release contracts, redaction maps, and irreversibility disclosures for SANITIZED/PUBLIC releases
• bounded unlearning contracts and post-unlearning verification receipts

Key idea:
Do not say:

*“we deleted it, so it is forgotten.”*

Say:

*“this subject was handled under this retention policy, propagation index, adjudication path, remedy contract, tombstone, semantic preservation receipt, extraction exclusion receipt, and bounded public claim.”*

Forgetting is not a button.

It is governance with receipts.

reacted to SeaWolf-AI's post with 🔥 1 day ago

Post

4795

Darwin V9 — GPQA Diamond 90.9%, #1 on the leaderboard, with pure greedy decoding
Darwin-398B-JGOS reaches 90.9% (180/198) on GPQA Diamond, the PhD-level scientific reasoning benchmark, ranking #1 on the Hugging Face GPQA Diamond leaderboard. No self-consistency, no test-time compute scaling — this was achieved with a single greedy decode (temperature 0, single sample, max 16,384 tokens). The full eval config is published in the model card, so anyone can reproduce it. Raw reasoning, no score inflation.
The result comes from Darwin V9, a patented evolutionary model-development platform. Its core idea: it never trains a model from scratch.
Why Darwin V9 beats training from scratch

Cost & speed: no trillion-token pretraining run, no months of compute — a purpose-built, high-performance model is produced in a fraction of the time.
Reuse of proven intelligence: instead of re-learning every capability from a blank slate, it selects and combines only the strengths of already-trained, already-validated models, so results are stable and predictable.
Surgical transplantation: it identifies which neural region of which model holds which capability — at the FFN (Feed Forward Network) layer level — and grafts in only the segments that contribute to the target skill.

How it works: a large model (Qwen 3.5 397B) serves as the mother model (the substrate); several father models specialized in reasoning, coding, and language are analyzed layer-by-layer across their FFN regions; the segments that contribute to the target performance are extracted and transplanted into the mother model to produce a new child model. The result is a ~400B MoE that activates only ~17B parameters per token at inference — large-model capacity with efficient inference.
If training from scratch means rebuilding everything from a blank page, Darwin V9 means precisely recombining intelligence that has already been proven. GPQA Diamond #1 is the proof.
Model: FINAL-Bench/Darwin-398B-JGOS

posted an update 2 days ago

Post

Here's my Hackathon Entries:
build-small-hackathon/ModuleMind
build-small-hackathon/neural-n64
build-small-hackathon/neural-riscv
build-small-hackathon/neural-gb
build-small-hackathon/neural-doom
build-small-hackathon/Flight-Transit-Agent
build-small-hackathon/HuggingWizards

I released a pretrained model base for the end of the Small Models, Big Adventures Hackathon. It is an 86M parameters model, trained on FineWebEdu. Somewhat customized from the standard model. I used my remaining Modal credits from the Hackathon

Model Weights + Demo:
Quazim0t0/Escarda-86M
build-small-hackathon/Escarda-86M-Chat

reacted to merve's post with ❤️❤️ 11 months ago

Post

2503

past week had huuuge releases 💗
here's our picks 🔥 find more models, datasets, demos here merve/releases-july-11-68750452c358c98b0fa663f7

> moonshotai/Kimi-K2-Instruct is the new sota LLM with 1T total 32B active parameters 🤯

> HuggingFaceTB/SmolLM3-3B is the new best LM for it's size, offers thinking mode 💭 as well as the dataset HuggingFaceTB/smoltalk2

> Alibaba-NLP/WebSailor-3B is the new agentic LLM for complex browsing

> Google DeepMind released medical vision LMs with an agentic doctor-patient app google/medgemma-release-680aade845f90bec6a3f60c4

> fal released a LoRA to improve details on face images fal/Realism-Detailer-Kontext-Dev-LoRA

reacted to CultriX's post with 🔥 11 months ago

Post

2382

New Space: Generate Knowledge Graphs from input data using LLM's (OpenRouter). It's a trial project but seems to be working alright so far!

CultriX/Generate-Knowledge-Graphs

Below is an example after feeding it the wikipedia page about Elon Musk:

reacted to lbourdois's post with ❤️ about 1 year ago

Post

3716

We introduce FAT5 (Flash Attention T5) ⚡

An implementation of T5 in PyTorch with UL2 objective optimized for GPGPU for both training and inference thanks to 13 different optimizations.
The main one is that we have designed a CUDA kernel to expand the Flash Attention by @tridao with RPE biases and supports other PE such as RoPE, ALiBi or FIRE.
The result kernel is 2 times faster than a SPDA implementation.
We also use Triton kernels to optimize certain parts of the architecture, such as the cross-entropy and RMSNorm layer.

The various kernels have been carefully built to be compatible with BF16 and torch.compile to go even faster and achieve efficient pretraining.

All other optimizations are described in a 📝 subsequent blog post available on @huggingface 🤗: CATIE-AQ/FAT5-report.

This methodology enabled us to efficiently pretrain as a proof of concept a FAT5 with 147M parameters in French in a reasonable time (1,461H for 419B tokens), with limited resources (1 A100 i.e. a computational budget of ~ €1,900) and a low carbon footprint (13.5kg eq CO2).

The model's weights are also available on Hugging Face: CATIE-AQ/FAT5-small.
Not very useful in practice, it's a PoC and not an instructed model (it's planned for later).

All the code is available on GitHub if you want to pretrain your own model in your own language or for a specific domain: https://github.com/catie-aq/flashT5 ⭐

Ending by indicating that was a joint project with @BorisAlbar at hf.co/CATIE-AQ.

reacted to AtAndDev's post with 👍 over 1 year ago

Post

4392

There seems to multiple paid apps shared here that are based on models on hf, but some ppl sell their wrappers as "products" and promote them here. For a long time, hf was the best and only platform to do oss model stuff but with the recent AI website builders anyone can create a product (really crappy ones btw) and try to sell it with no contribution to oss stuff. Please dont do this, or try finetuning the models you use...
Sorry for filling yall feed with this bs but yk...

6 replies

reacted to burtenshaw's post with 🤗 over 1 year ago

Post

2105

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

Dean Byrne PRO

AI & ML interests

Recent Activity

Organizations

Quazim0t0's activity