Yusuf Chowdury

Yusufchy

https://aipulsedaily.vionixai.tech/

AI & ML interests

AI agents, open-weight models, machine learning, MLOps, developer tools, AI automation, and AI-assisted publishing.

Recent Activity

repliedto Banaxi-Tech's post 28 minutes ago

Today we wanted to release BananaMind 2 Pico, our smallest model yet at ~0.9M parameters. Instead, we accidentally ran a very expensive experiment on what happens when you push a tiny model way past its useful token budget. Short version: we trained on 200B tokens (~222K:1 tokens-per-parameter). The model peaked at 20B tokens with an INT Index of 4.55, then degraded monotonically over the next 160B to 3.31 — a 27% regression. Three of four Open SLM benchmarks were worse at the end of training than they were at 10% through. The useful compute-optimal range for Pico-tier models looks like ~22K–30K tokens per parameter. Ratios like 7K:1, 15K:1, and 22K:1 all work fine — TinyStories and most sub-3M community models sit in this range. Push much further and benchmarks start rotting. Follow us for more: https://huggingface.co/BananaMind @vovaRL @Banaxi-Tech Full writeup with all checkpoints, the Chinchilla-ratio control run, and the schedule-vs-overtraining analysis: https://huggingface.co/blog/Banaxi-Tech/ovdadadadd And if anyone, i dont know the reason why you would, wants the 20B token checkpoint reply and ill upload it as BananaMind 2.1 Pico EXP

repliedto RDTvlokip's post about 5 hours ago

I published a second article about a reader taking apart four of my numbers. He came back a fifth time, and went to the experiment that had never run. He found the one derived threshold in it was built on a sample maximum. He was right, and there was worse. That maximum cannot converge, because the codes I wanted to declare unreachable are themselves inside the null distribution. Its supremum is exactly the value I was excluding. So I ran the whole thing. Seven questions in one day, in a world of 27 referents small enough that the optimum, the null distribution and the gradient are computed rather than estimated. The certificate carrying the project does not survive two agents. A symmetry argument replaces it, with a corollary I did not expect: a free per-object embedding table cancels in advance anything the message structure could contribute. No training trick recovers it. You can check that by reading an architecture, without running it. Eight of my hypotheses died that day. Not one was an arithmetic error. Then I did the literature review, last, and found the argument published in 2021. The real lesson: Shrinking a world until everything is exact removes one class of mistake and leaves untouched the class that was doing the damage. Twenty minutes of searching would have saved a day of deriving. What caught my errors was never the exactness. It was an outside reader, a second route to the same number, and predictions written down before measuring. Code, every number including the ones I would have cut, and sixteen dated refutations 👇 🔗 https://huggingface.co/blog/RDTvlokip/i-made-my-world-small-enough-to-compute-everything 💻 https://github.com/RDTvlokip/RDTRL 📦 https://doi.org/10.5281/zenodo.21726216

repliedto dronefreak's post about 5 hours ago

🚀 Excited to open-source the SeaDronesSee Object Detection Model Zoo on Hugging Face. This release includes: - 🤖 YOLOv8, YOLOv11, YOLOv26 and RF-DETR object detection models trained on SeaDronesSee, spanning nano through x-large YOLO variants plus RF-DETR Nano/Small/Medium. - 🌊 Benchmarked on SeaDronesSee's maritime search-and-rescue setting — swimmers, boats, jet skis, life-saving appliances and buoys captured by UAVs over open water, at varying altitudes and non-uniform image resolutions (1080p up to 4K+). - 📊 Detailed model cards with mAP/precision/recall, per-class breakdowns, PR/F1 curves and confusion matrices (YOLO), qualitative detection showcases, and full training configurations for reproducibility. Headline numbers: - 🏆 Best mAP@50: 83.47% (RF-DETR Medium), 47.49% mAP@50:95, 87.01% precision. - ⚡ Best efficiency tradeoff: YOLOv26s hits 80.14% mAP@50 at just 22.8 GFLOPs (10.0M params) — within ~3 points of the top RF-DETR variant, while actually beating YOLOv11x's 74.82% mAP@50 using ~8.6x fewer FLOPs (196.0 GFLOPs). The goal is to make benchmarking and experimenting with maritime UAV perception easier by providing ready-to-use pretrained checkpoints, all trained and evaluated under one shared pipeline (DetectionBench: https://github.com/dronefreak/DetectionBench). Full credit for the underlying dataset goes to Leon Amadeus Varga, Benjamin Kiefer, Martin Messmer, and Andreas Zell (University of Tübingen, WACV 2022) — this release is an unofficial, YOLO-ready reformatting of their work (CC0-licensed), not a new dataset. If you're working on maritime search-and-rescue, UAV perception, autonomous drones, or real-time object detection, I hope these resources are useful. 📦 Dataset: dronefreak/SeaDronesSee 🤖 Model Collection: https://huggingface.co/collections/dronefreak/seadronessee-object-detection-model-zoo-6a7b030a25797e5dd2d70123 Feedback, bug reports, and contributions are always welcome.

View all activity

Organizations

None yet

Yusufchy 's models

None public yet