👋 Open to Work

Joseph [open/acc] Pollack PRO

Tonic

hugging-science

·

https://discord.gg/qdfnvSPcqP

AI & ML interests

🤖Making robots to help people learn things quicker 👩🏻‍🚀🚀

Recent Activity

reacted to RDTvlokip's post with 👍 2 days ago

I finally changed the architecture of my 15M French LLM. It worked. Then I almost fooled myself about how much and catching that was the real win. After proving last time that architecture is a threshold, not a lever, I got stubborn: could I change how the model learns? Four honest attempts, Lion, a sharper AdamW β2, multi-token prediction, LayerScale. Four failures. The bottleneck wasn't the learning rule either. So I changed the shape of the computation instead: loop the same transformer blocks 4×, deeper reasoning, zero added parameters. It beat the baseline on perplexity, the first thing in the whole project to move that number. Then I added my own twist: let each token decide how deep to think, halting on its own entropy. My first evaluation was spectacular. Coherence up 65%. Hallucinated names down 62%. It was noise. Eight prompts, one seed. I re-ran on 50 prompts × 200 tokens and watched the gains shrink to "modest" and on out-of-domain prompts, recurrence actually made things worse. No universal winner. And none of it is new: it's Adaptive Computation Time (2016), the Universal Transformer (2018), and LoopViT (2026), recombined and measured honestly. The real lesson: A number from 8 prompts is a rumor. The eval harness that kills your own best result is worth more than the result it kills. Cite your lineage. Stay preliminary until multiple seeds say otherwise. The three models are live. The write-up is honest about every caveat 👇 🔗 https://huggingface.co/blog/RDTvlokip/teaching-a-15m-french-llm-to-think-deeper

liked a Space 4 days ago

liked a Space 7 days ago

julien-c/caliceo

View all activity

Organizations

Tonic 's Spaces 165

G Android Control

Visualize experiment metrics with interactive line charts

L Android Control

Visualize training run metrics in real time

Web Search MCP

Search and extract web content for LLM ingestion

SmolFactory

SmolFactory is a e2e model maker

Nanonets Ocr S

MCP server For OCR made with Gradio by Nanonets

Track Tonic

trackio for training monitoring

OpenReasoning Nemotron 14B

NemoAligner Synthetic SFT with function calling

Openai Gpt Oss 20b

Generate text using a powerful language model

Gemlite Llama3.1

seriously ?? 5.4GB RAM gemlite quantized llama3.1

tonic-testing-paywalls-app-18666

Moonshot Math

Formal reasoning model that can reason and prove theorems

Tonic's Lucie 7B

A french-speaking LLM trained with open data

Tonic's GOT OCR

GOT - OCR (from : UCAS, Beijing)

Yi Coder 9B

state-of-the-art coding performance with fewer than 10 B

Leaderboard Yourbench Tonic ESMA-Auto-Bench

Display leaderboard and analyze samples

Native 1-bit LLM

1.58 Bits Native Micro LLM from Microsoft

Command A

OpenStreetMap AI Helper

Contributing to OpenStreetMap with the help of AI

Llava Video

interact with videos !

Falcon Vision

Falcon2-11B-vlm is a vision model by TIL

StarCoder2

Latest Coding Model By🤗Huggingface & Friends-101 languages!

First Agent Template

Get the current time in any timezone

Submission Template

Classify text into categories

Phi 4

Microsoft's iconic Phi-4 model