Flora Chen

Most ML models don’t fail because of bad algorithms. They fail because the right questions were never asked. Before you open a notebook, choose a model, or tune hyperparameters - there are 5 questions that decide whether your model will ever be used. Save this checklist. Share it with your team. And build models that actually get used. #DataDerivators #DataScienceCommunity #DataScience #MachineLearning #AppliedAI #DecisionIntelligence #RealWorldAI #AI

2

Like recall and precision metrics in ML, we have been diving into fidelity and privacy metrics for synthetic data generation. As we explore privacy-safe data solutions at Rockfish, understanding these trade-offs is becoming crucial. If you're curious about the technical details, this is a must-read!

6

There is Bad AI. And then there is Chiron. Not all AI is created equal. Bad AI: Query time reasoning on raw telemetry. Latency compounds at scale. Black box inference with no causal trace engineers can inspect. Alert volume stays high. Correlation happens after the fact. Infrastructure costs scale linearly with data volume with no ceiling. Then there is Chiron: Causal rules compiled at build time. This delivers sub 500ms deterministic RCA for the fastest possible MTTR. Single actionable alert per incident. Causality is determined before impact. Every alert carries a full evidence chain back to the root cause. Query store free architecture. A streaming state store runs alongside existing infra with no rip and replace. Compressed state over raw logs. The cost profile is decoupled from data volume. SREs have been trained to react faster. The better engineering problem is asking why we are reacting at all. That is the architectural shift Chiron is built on.

12

2 Comments

When your compute budget is limited, should you prioritize model parameters or training tokens? Shuyang Xiang's experiment with tiny transformers shows that for smaller budgets, smaller models trained on more data often perform better.

3

Prior Labs just dropped TabPFN-2.5, a tabular foundation model that learns from 50,000 samples and 2,000 features without training, matching AutoGluon! Prior Labs releases TabPFN-2.5, a tabular foundation model scaling in-context learning to 50,000 samples and 2,000 features without training. It outperforms tuned tree-based models and matches AutoGluon 1.4 on benchmarks. Key features include alternating attention, synthetic task training, and a distillation engine for compact model deployment. * Scales tabular transformers to 50K samples and 2K features. * Outperforms tuned tree baselines and matches AutoGluon 1.4. * Enables permutation invariance and in-context learning. This development marks a significant leap in efficient, high-performance tabular data analysis. Originally by Asif Razzaq — "Prior Labs Releases TabPFN-2.5: The Latest Version of TabPFN that Unlocks Scale and Speed for Tabular Foundation Models" (Sat, 08 Nov 2025 10:44:44 +0000) — https://lnkd.in/eCJj499U How might TabPFN-2.5's ability to scale in-context learning without training impact future tabular data analysis workflows? #TabPFN #tabularData #machineLearning #foundationModels

5

A lightweight VLM outperforms much larger models in OCR. With just 1.7B parameters, dots.ocr sets a new state of the art on OmniDocBench, showing that strong document understanding doesn’t require massive models or costly APIs. Most current OCR leaders rely on very large models (72B+) or paid services like GPT-4o. dots.ocr takes a different approach. Built on a compact 1.7B LLM foundation, it outperforms heavyweight models such as Qwen2-VL-72B and GPT-4o across key benchmarks. Why it stands out: ➡️ Unified architecture: Processes text, tables, formulas, and layout in a single pass ➡️ High precision: Achieves an overall edit distance of 0.125 on ➡️ OmniDocBench (lower is better), ahead of GPT-4o at 0.233 ➡️ Production-ready: Integrated with vLLM for high-throughput deployment Performance highlights: ➡️ Table recognition: 88.6 TEDS score (vs. 72.0 for GPT-4o) ➡️ Layout detection: 0.930 F1 score, outperforming DocLayout-YOLO ➡️ Inference speed: Faster than larger foundation models thanks to its compact size You can deploy dots.ocr locally using vLLM or Hugging Face. The GitHub repo and huggingface page are in the comments.

71

3 Comments