In a Training Loop 🔄

John Smith PRO

John6666

John6666cat

AI & ML interests

None yet

Recent Activity

liked a model about 8 hours ago

prithivMLmods/Qwen3-VL-8B-Heretic-Stable

liked a model about 8 hours ago

mradermacher/Qwen3-VL-8B-Heretic-Stable-i1-GGUF

liked a model about 8 hours ago

mradermacher/Qwen3-VL-8B-Heretic-Stable-GGUF

View all activity

Organizations

reacted to Jiaqi-hkust's post with 👀 1 day ago

Post

2364

Happy to introduce Response-G1 #ACL2026 — a proactive agent for streaming video understanding.

📄 Paper: http://arxiv.org/abs/2605.07575
📷 Code: http://github.com/kadmkbl/Response-G1

We are happy to have a further discussion!!!

#ACL2026 #AI #Multimodal #VideoUnderstanding #OpenSource #LLM

1 reply

reacted to danielhanchen's post with 👍🤗 1 day ago

Post

2690

Google releases Gemma 4 QAT. ✨
You can now run Gemma 4 at 3x less memory with near original performance.

QAT makes it possible to run Gemma 4 26B-A4B on 16GB RAM.

GGUFs: https://huggingface.co/collections/unsloth/gemma-4-qat
QAT Guide: https://unsloth.ai/docs/models/gemma-4/qat

reacted to pbhappliedsystems's post with 👀 1 day ago

Post

1646

🚀 **New flagship dataset — and an argument about what a dataset card should be.**

Most synthetic datasets on the Hub ship row counts, a license, and little else — pipeline opaque, rejection criteria unstated, compliance unaudited. We published the opposite.

**SynthEval Cloud — Regulated-Domain Synthetic Instruction Dataset**
👉 pbhappliedsystems/syntheval-cloud-regulated-instruct-1k

**1,116** quality-gated instruction records across **7 regulated domains** (medical, legal, GDPR, privacy, education, e-commerce, transport). Every record cleared a documented cascade, not a vibe check:

- 🧪 **Dual-signal hallucination gate** — rejects only when embedding cosine *and* keyword-overlap both fail; a low score alone never rejects.
- 🔒 **Layered PII masking + independent leak audit** — a separate over-reporting scanner found **0.0% residual leak** across all 1,116 records.
- 📊 **Whole-corpus evaluation, not a sample** — MATTR **0.769**, mean cosine **0.73**, **0%** near-duplicates, **96.9%** yield.
- 🧾 **The 36 rejections ship too**, each tagged with its failing gate. Removal at the gate is the product; we show our work.

Every number on the card is a field in the evaluation_report.json shipped beside the data — full methodology + provenance (Mistral-Nemo AWQ W4A16 · vLLM 0.8.5.post1 · Modal A10G).

One release from **SynthEval**: Studio (local GPU) + Cloud (Modal+vLLM), proving quality parity across substrates.

📄 Whitepaper: https://pbhappliedsystems.com/SynthEval_Studio_and_Cloud_Quality-Gated_Synthetic_Data_Generation.pdf
🔎 Overview: https://pbhappliedsystems.com/synthetic-data.html

**CC BY 4.0** — commercial use welcome, just credit it. Need defensible synthetic data at scale? Let's talk.

— Patrick Hill, PBH Applied Systems

reacted to azettl's post with 👍 1 day ago

Post

883

𝗕𝘂𝗶𝗹𝗱 𝗦𝗺𝗮𝗹𝗹 𝗛𝗮𝗰𝗸𝗮𝘁𝗵𝗼𝗻 - 𝗪𝗲𝗲𝗸 𝗶𝗻 𝗣𝗿𝗼𝗴𝗿𝗲𝘀𝘀

After Consilium, I wanted to build something entirely different. Smaller. Weirder.

The idea came from a single Asterix scene—Asterix and Obelix trying to obtain Permit A38, only to discover you need Permit A38 to apply for Permit A38. That's the whole game.

𝗪𝗵𝗮𝘁 𝗜 𝗯𝘂𝗶𝗹𝘁: A bureaucratic text adventure powered by one fine-tuned 1.7B model playing five different characters—each with their own system prompt, their own rules, and their own way of sending you in circles.

𝗧𝗵𝗲 𝗳𝗶𝘃𝗲 𝗽𝗲𝗿𝘀𝗼𝗻𝗮𝗹𝗶𝘁𝗶𝗲𝘀:

✅ Clerk Vitalstatistix—requires Form 27b/6. Has never issued a permit in 23 years.

✅ Supervisor Caligula Minus—perpetually at lunch. Invented A38 in 1987, forgot why.

✅ SYSTEMA v2.3—last updated 1994. Every response starts with an error code.

✅ Form 27b/6—sent. Not happy about it. Page 3 is always missing.

✅ Ombudsman Panoramix—investigates complaints about the process. It is also the process.

𝗧𝗵𝗲 𝘁𝗲𝗰𝗵:

- Fine-tuned SmolLM2-1.7B on ~1000 synthetic NPC examples generated with Claude.
- Converted to GGUF Q4_K_M for fast CPU inference
- Streaming responses so you feel the bureaucracy arriving in real time
- Custom beige government office UI

➡️ Try it: azettl/the-place-that-sends-you-mad

Built for the Build Small Hackathon—Track 2: Thousand Token Wood. Model at azettl/permit-a38-npc and Dataset azettl/permit-a38-npcs.

reacted to Sathya77's post with 🔥 1 day ago

Post

520

Built a ViT for ×4 image super-resolution from scratch in PyTorch — sharing the model.

No pretrained weights. Every component implemented from scratch: strided Conv2d patch embedding, multi-head self-attention across 1,024 tokens, 6 pre-norm transformer blocks, and a PixelShuffle reconstruction head for learned upsampling.

Trained on real-images from LSDIR dataset with fp16 AMP on a laptop GPU. Tiled inference handles arbitrary input sizes.

Current architecture: patch size 2, embed dim 64, 4 attention heads, 6 transformer blocks, ~786K parameters — test PSNR 23.30 dB.

The model handles broad structure well — fine textures and sharp edges need more capacity. Working on a larger configuration next.

🤗 Space: Sathya77/ViT-ISR-Tiny-LSDIR

Feedback welcome — especially on the architecture choices.

reacted to mmhamdy's post with 🚀 1 day ago

Post

958

Human brains don't recreate every pixel to understand the world!

Most current models in genomics, proteomics, and single-cell transcriptomics rely on generative objectives like masked language modeling or next token prediction. While effective, these architectures waste significant capacity reconstructing raw, noisy sequence details that may not carry functional biological meaning.

But a promising, more efficient alternative is emerging: Joint-Embedding Predictive Architecture (JEPA)

Originally introduced by Yann LeCun for computer vision, JEPA is a non-generative, self-supervised learning (SSL) framework. Instead of predicting raw inputs, it operates as a world model that predicts abstract semantic embeddings in latent space.

Recently, the JEPA framework (and its more efficient LeJEPA variant) has been adapted into the biological sciences to develop performing foundation models and to improve on already existing ones.

It's interesting how each adaptation modified and tailored JEPA to suit its specific biological domain, whether by experimenting with different backbones or complementing the objective with other loss terms.

For example, JEPA-DNA and ProteinJEPA used JEPA as a continual pre-training framework to enhance existing foundation models without training from scratch, while Cell-JEPA and JEPA-DNA employed a hybrid objective that combines the JEPA loss with a traditional language modeling loss.

The article below provides an overview of these implementations, along with others that came out this year. As always, your thoughts and feedback are welcome and highly appreciated!

Link to the article is in the first comment 👇

3 replies

reacted to dippatel1994's post with 🔥 1 day ago

Post

956

To make revising LLM architectures and training methods faster, I created a deck of 180 visual flashcards. It started as a personal hobby, but slowly became cheat code for reviewing LLM concepts before technical interviews. People love it!

Swipe through these samples, and if you want to grab the full set or follow the project, the repo is here: https://github.com/llmsresearch/llm-flashcards.

reacted to matteospanio's post with ➕🚀 1 day ago

Post

6966

🎶 Released mule-torch — an unofficial PyTorch port of MULE (SF-NFNet-F0), SiriusXM/Pandora's music-audio embedding model (McCallum et al., ISMIR 2022).

No retraining: I re-implemented the architecture in pure PyTorch and transferred the original TensorFlow weights, then checked it layer by layer against the genuine TF pipeline.

✅ End-to-end clip-embedding cosine 0.9999999 vs the original
✅ ONNX backbone parity < 1e-6
✅ 62.35M params (paper: ~62.4M)
✅ Batched, GPU-native, ONNX-exportable — none of which the original Analysis pipeline does

pip install mule-torch

from mule_torch import MuleModel
emb = MuleModel.from_pretrained()(waveform)   # (B, T)@16kHz -> (B, 1728)

🤗 Weights: matteospanio/mule
💻 Code: https://github.com/matteospanio/mule-torch
📦 PyPI: https://pypi.org/project/mule-torch/

The fun bug: parity was perfect through every conv but the block output was anti-correlated (cos = −1). Cause: the learnable skip-init gains couldn't be mapped by layer name (Keras scrambles the order) — they had to be recovered from the graph.

⚠️ Unofficial, community port — not affiliated with or endorsed by the original authors. All credit to them; please cite the paper. Weights inherit CC-BY-NC-4.0.

reacted to TensorVizion's post with 👀 1 day ago

Post

Check Out My Structured Data Gradio Space for the Build Small Hackathon. Short Demo Video. build-small-hackathon/Structured-Data-Rescuer

reacted to pankajpandey-dev's post with 🔥 1 day ago

Post

855

🇮🇳 Gemma-3-1B Hindi Instruct — a Hindi LLM that runs fully offline, anywhere.
Last week I shipped Qwen3-4B Hindi. This week I went the other direction: how tiny can a useful Hindi model get? So I fine-tuned Gemma-3-1B on quality-filtered Hindi instruction data and shipped the full GGUF ladder.
✅ Fine-tune (16-bit): pankajpandey-dev/gemma-3-1b-hindi-instruct
✅ GGUF (Q4/Q5/Q8): pankajpandey-dev/gemma-3-1b-hindi-instruct-GGUF
Runs in Ollama, llama.cpp, and LM Studio. The Q4_K_M is just 806 MB — runs on CPU, a cheap laptop, even a Raspberry Pi.
What I tried this round: chrF-filtered the training data to drop weak translations, and used response-only loss so the model learns how to answer, not how to repeat prompts.
Honest note: at 1B, Hindi fluency is strong but coherence is bounded by size — it's a lightweight/edge experiment, not a 4B replacement. Gemma-3-4B Hindi is next.
Part of my Hindi LLM Series — openly-licensed Indic models for local & edge use. Feedback welcome 🙏
#Hindi #IndicNLP #GGUF #LocalLLM #Gemma #EdgeAI

reacted to kanaria007's post with 👀 1 day ago

Post

156

✅ Article highlight: *Interop Schemas for Learning-World Governance Artifacts* (art-60-175, v0.1)

TL;DR:
This article argues that governance without interop is vendor-local theater.

It is not enough for one system to say *“we have receipts.”* If another vendor cannot parse the artifact, reproduce the digest, replay the bundle, and reach the same admissibility outcome, the claim is not really portable. So 175 defines a common interop layer: shared envelopes, pinned canonicalization, minimal portable schemas, and deterministic bundle formats.

Read:
kanaria007/agi-structural-intelligence-protocols

Why it matters:
• turns governance artifacts into cross-vendor verifiable objects rather than local implementation details
• fixes the classic failure modes of digest drift, schema drift, and bundle drift
• makes “same artifact / same verdict” a testable claim instead of a handshake promise
• gives courts, forgetting flows, and unlearning claims portable bundle formats

What’s inside:
• a common *interop envelope* for contracts, manifests, receipts, and bundles
• a pinned *canonicalization profile* plus conformance receipts to stop digest disagreements
• minimal portable schemas for core learning-world governance artifacts
• deterministic bundle formats like *Court ZIP*, *Forgetting ZIP*, and *Unlearning ZIP*
• replay/conformance receipts so another vendor can verify the same bundle and reach the same admissibility result

Key idea:
Do not say:

*“our system can export the evidence.”*

Say:

*“this artifact uses this schema registry, this canonicalization profile, this interop-safe digest model, and this bundle index—so another vendor can verify the same object and reach the same result.”*

That is how governance stops being local theater and becomes portable infrastructure.

reacted to marinarosa's post with 🔥 1 day ago

Post

I'm participating in the Build Small Hackathon! 💖

I chose the Backyard AI chapter: solve a real problem for someone you know.

My idea is to help a business owner extract valuable information from their data. She operates, negotiates and provides support for customers all through Whatsapp. She doesn't use Notion, Obsidian, Spreadsheets and does not like doing repetitive data-entry tasks. It's a messy, chaotic messaging workflow that kind of works for her now.

What if she could open a Space, upload her exported WhatsApp data, and a fine-tuned model tailored to her business domain and conversation style extracted, classified, organized all her customers and deals into a coherent dashboard, along with a chatbot for her to ask questions about her business? That's my current approach.

5 replies

reacted to sergiopaniego's post with 🧠 1 day ago

Post

207

Frontier agents are this good partly because the model was trained inside the very harness it ships with.

NVIDIA's new paper "Polar: Agentic RL on Any Harness at Scale" brings that recipe to the open: it turns coding harnesses like Codex, Claude Code, Qwen Code or Pi into RL training environments without touching their internals.

The core idea: every agent, however complex or closed, talks to a model through an API, so they put a proxy there. The harness runs exactly like in production while the proxy records prompts, sampled token ids and logprobs. Trajectories get rebuilt outside, token faithful, so gradients hit the exact tokens the policy sampled.

The gains are consistent across all four harnesses. Same Qwen3.5-4B, plain GRPO, evaluated on SWE-Bench Verified:

Codex 3.8 → 26.4 (+22.6)
Claude Code 29.8 → 34.6 (+4.8)
Qwen Code 34.6 → 35.2 (+0.6)
Pi 34.2 → 40.4 (+6.2)

The biggest gains appear on unfamiliar execution paths, Codex being the clearest case. The takeaway: you are not just training a model, you are training the model + harness system.

Two engineering pieces make it work at scale. Async worker pools isolate container boots (CPU), agent execution (GPU) and long tail test runs, so slow runtimes never block the GPUs. And prefix merging stitches hundreds of captured API calls back into contiguous traces: 5.4x faster trainer updates and rollout GPUs at 88% utilization.

It also doubles as an SFT data factory: 504 test verified agent traces from a 122B teacher, multi-turn conversations averaging 104 messages each, coming to the Hub under Apache 2.0 (release pending review).

Paper authors: Binfeng Xu, Hao Zhang, Shaokun Zhang, Songyang Han, Mingjie Liu, Jian Hu, Shizhe Diao, Zhenghui Jin, Yunheng Zou, Michael Demoret, Jan Kautz and Yi Dong.

> Paper: Polar: Agentic RL on Any Harness at Scale (2605.24220)
> Code: https://github.com/NVIDIA-NeMo/ProRL-Agent-Server
> Training data: NovaSky-AI/SkyRL-v0-293-data

reacted to TravisMuhlestein's post with 🧠 1 day ago

Post

One thing that stands out to me about efforts like DNS-AID and ANS is a simple architectural principle:

We didn't reinvent the internet. We extended it.

A lot of discussion around agent ecosystems focuses on models, orchestration frameworks, and capabilities. But once agents begin operating across platforms and organizational boundaries, a different set of questions emerges:

• How are agents discovered?
• How is identity established?
• How is trust verified?
• How are capabilities described in a machine-readable way?

DNS-AID and ANS approach different parts of the same problem space using infrastructure that already exists and operates at internet scale.

• Identity standards help establish trust

• Discovery standards help locate capabilities and services

One thing I'm increasingly convinced of: we have model cards, but we still lack broadly adopted capability manifests for agents.

As agent ecosystems evolve, open standards around discovery, identity, trust, and capability description may become just as important as the models themselves.

Curious what builders here think: what's the biggest missing standard today for truly interoperable agent systems?

reacted to appvoid's post with 👀 1 day ago

Post

119

yikes! i missed the small model hackathon i guess i'll have to make sota for people to notice

reacted to OpceanAI's post with 🧠 1 day ago

Post

**Two Years of OpceanAI**

Building AI infrastructure without funding should not be possible. We have been doing it for two years.

OpceanAI was founded in June 2024 as a personal bot project. Today it is an ecosystem: Doki, the container engine for Android; ToS, the P2P data movement protocol; Ixari, the 140GB corpus spanning Spanish and ten indigenous languages; and Yuuki, the model family built without cloud infrastructure. Everything open source. Everything free. Everything built on a single conviction: the barrier is mindset, not money.

OpceanAI has no external funding. That has not changed. It does not slow the work — it shapes it. Every component is designed to last, because we cannot afford to rebuild it.

**Two Models in Development**

**Yumo RxG · 32B**
A 32-billion-parameter language model. Yuuki architecture converged with Yumo's reasoning layer. Two systems unified into one, built for deep reasoning at scale.
Status: advanced development.
huggingface.co/YU-MO/Yumo-RxG

**Yuuki RxG VL · 27B**
A 27-billion-parameter multimodal model — text and vision — built on the Yuuki architecture. Yuuki has always prioritized honesty over performance theater. VL extends that property to everything the model perceives.
Status: advanced development.
huggingface.co/OpceanAI/Yuuki-RxG-vl

No release date is set. We announce them today because the work is real, and this anniversary is the right moment to mark where we are going.

**To Those Who Have Followed This**

Every bug report, every GitHub star, every download has been part of what comes next. This is built with you, not for you.

Thank you. We continue.

— The OpceanAI Team :]

reacted to ManniX-ITA's post with 👀 1 day ago

Post

114

🚀 Gemma-4-A4B 98e v7-coder cohort — two 20.8B MoE coders (4B-active), fresh-map prunes of Gemma 4 26B-A4B. 30/128 experts dropped per layer from the rebuilt v7 competence maps (audited producers, 10 classes), generic-code 3× + LiveCodeBench 2× on a [24,40] floor, plus the mandatory shared-FFN α=1.2.

📊 Q6_K, llama.cpp, greedy, same host (from summary.json):

🧪 v7-coder — science-augmented (+ targeted_gpqa 1.5). A coder that kept ALL its science: GPQA-D 70.71 (+9.6pp over v6-coder, at PARITY with the unpruned 128e's 67.17), AIME 76.67, MATH-500 92.0, GSM8K 93.0, HE 98.78, HE+ 92.68, LCB-55 96.36, LCB-100 97.0, MultiPL-E 88.67, IFEval 95, ARC 94.8. Edges 128e on GPQA/AIME/GSM8K with no code regression — science recovered by a dedicated targeted_gpqa calibration class.

⚡ v7-coderx — code-maximal (no science term). The strongest coder in the cohort: LCB-med-55 98.18 + LCB-100 99.0 (highest of any Gemma-4 prune to date, +1.8/+2.0pp past the unpruned 128e), MultiPL-E 90.0, HE+ 92.68, HE 95.73, IFEval 95, GSM8K 91, MATH-500 89, AIME 70.0. Whole budget on code; the trade is graduate science (GPQA 48.48).

🎯 v7-coder for strong code + graduate science; v7-coderx for max code (~1.8pp LCB-55 buys ~+22pp GPQA between them). Successor to v6-coder, whose code profile led the 14–22B band (+9pp HE over Qwen2.5-Coder-14B, same-rig).

📦 Each ships bf16 · GGUF (29 tiers + ContribDynamic CD-* + F16 + imatrix + mmproj vision) · NVFP4A16 (native vLLM, ~13 GB) · Ollama (29 tiers + vision-* + :latest=Q4_K_M).

🔗 v7-coder: ManniX-ITA/gemma-4-A4B-98e-v7-coder-it (+ -it-GGUF, -NVFP4A16) · https://ollama.com/mannix/gemma4-98e-v7-coder
🔗 v7-coderx: ManniX-ITA/gemma-4-A4B-98e-v7-coderx-it (+ -it-GGUF, -NVFP4A16) · https://ollama.com/mannix/gemma4-98e-v7-coderx

reacted to Hari5115's post with 🚀 1 day ago

Post

How do scientists know if a new chemical is toxic before testing it on anything alive ?

For decades: animal studies. Slow, expensive, ethically complicated.

Tox21 changed that — a government-backed initiative screening 10,000 compounds across 12 biochemical assays, testing whether molecules activate receptors linked to cancer, hormonal disruption, and organ damage.

I trained a model on this data and published everything under MIT — free for research, education, and building.

🔬 Hari5115/molecular-toxicity-predictor
📦 Hari5115/MoleculeIQ
🚀 Hari5115/molecular-toxicity-demo

⚠️ Research and educational use only. Not a substitute for certified toxicological testing. Model has known limitations on novel chemical classes — see model card for details.

What would you build with this? 👇

#OpenSource #DrugDiscovery #ChemInformatics #Tox21 #HuggingFace

John Smith PRO

AI & ML interests

Recent Activity

Organizations

John6666's activity