·
AI & ML interests
None yet
Recent Activity
posted an update 1 day 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 View all activity Organizations
view article 🔁 Apprendre à un LLM français de 15M à penser plus profond — et à savoir quand s'arrêter 🇫🇷
RDTvlokip
• • 1
view article 🔁 Teaching a 15M French LLM to think deeper — and to know when to stop 🇫🇷
RDTvlokip
• • 1
view article 🔧 L'architecture est un seuil, pas un levier — ce que j'ai appris en optimisant un LLM français de 15M de paramètres 🇫🇷
RDTvlokip
• • 1
view article 🔧 Architecture is a threshold, not a lever — what I learned optimizing a 15M French LLM 🇫🇷
RDTvlokip
• • 1
view article 🧠 I trained my own French LLM from scratch — alone, with a 1080 Ti, and the power went out ⚡🇫🇷
view article 🧠 J'ai entraîné mon propre LLM français from scratch — seul, avec une 1080 Ti, et le courant a coupé ⚡🇫🇷
view article 🧲 Embeddings — When AI turns words into GPS coordinates! 📍🧠
view article 🧲 Embeddings — Quand l'IA transforme les mots en coordonnées GPS ! 📍🧠
view article 🎯 PCA (Principal Component Analysis) — Compresser les dimensions comme un boss ! 📊🔥
view article 🎯 PCA (Principal Component Analysis) — Compressing dimensions like a boss! 📊🔥
view article 🎯 K-Means — Quand l'IA organise le chaos en boîtes bien rangées ! 📦✨
view article 🎯 K-Means — When AI organizes chaos into neat boxes! 📦✨
view article 🎯 F1-Score — Quand l'Accuracy te ment en pleine face ! 📊💥
view article 🎯 F1-Score — When Accuracy lies to your face! 📊💥
view article 🎯 Precision & Recall — Les métriques jumelles qui ne sont jamais d'accord ! ⚖️🔍
view article 🎯 Precision & Recall — The twin metrics that never agree! ⚖️🔍
view article 🎆 AI 2026 — The 9 trends that will EXPLODE this year! 🚀💥
view article 🎆 IA 2026 — Les 9 tendances qui vont exploser cette année ! 🚀💥
view article 📊 Cross-Entropy — The loss function that KNOWS how to punish! 🎯🔥
RDTvlokip
• • 1
view article 📊 Cross-Entropy — La fonction de perte qui SAIT punir ! 🎯🔥
RDTvlokip
• • 2