koscak.ai Research Blog https://koscak.ai/blog/ Methods, run evidence and honest caveats from the koscak.ai research lab: KSS-LoRA training, NOESIS reasoning, weather, security. en Zero Bluffed Answers: How NOESIS Is Designed to Never Make Things Up https://koscak.ai/blog/noesis-zero-bluff-ai-design/ https://koscak.ai/blog/noesis-zero-bluff-ai-design/ Verify or abstain, running live: a zero-LLM engine that answers only what it can prove. One Sensor vs Three Global Models: Beating ICON, GFS and ECMWF https://koscak.ai/blog/hyperlocal-weather-beats-global-models/ https://koscak.ai/blog/hyperlocal-weather-beats-global-models/ One station, nine years of data, one Raspberry Pi - measured on held-out 2026 data. Your Security Team Is Guessing: Why Findings Must Be Chained, Not Listed https://koscak.ai/blog/security-findings-chained-not-listed/ https://koscak.ai/blog/security-findings-chained-not-listed/ Chained-finding analysis under exact-or-abstain reporting - starting with a free scan. Negative Overfitting Gap: The Koščák Coefficient and What It Means for LLM Fine-Tuning https://koscak.ai/blog/negative-overfitting-gap-koscak-coefficient/ https://koscak.ai/blog/negative-overfitting-gap-koscak-coefficient/ Validation loss below training loss - the model generalises better than it memorises. First confirmed instance in LLM fine-tuning. κ = −0.009. What it means and how to reproduce it. Why FP8 Training Silently Destroys Your Model Quality https://koscak.ai/blog/fp8-training-quality-loss/ https://koscak.ai/blog/fp8-training-quality-loss/ Standard LoRA at FP8 loses 68% quality without any warning signs. Here's the root cause and the fix. Antifragile AI: Why KSS-LoRA Gets Better With Noisy Training Data https://koscak.ai/blog/antifragile-ai-noise-robust-fine-tuning/ https://koscak.ai/blog/antifragile-ai-noise-robust-fine-tuning/ 50% noise injection reduces the overfitting gap by 94.7%. Most methods break under noise. This one uses it as fuel. LoRA Fine-Tuning on NVIDIA B300 Blackwell Ultra: Full Guide https://koscak.ai/blog/lora-fine-tuning-b300-blackwell/ https://koscak.ai/blog/lora-fine-tuning-b300-blackwell/ Complete guide to fine-tuning on B300 Blackwell Ultra with FP4 precision. Setup, config, and results. What Is Stochastic Sparse LoRA and Why Does It Work? https://koscak.ai/blog/stochastic-sparse-lora-explained/ https://koscak.ai/blog/stochastic-sparse-lora-explained/ The Bernoulli sparsity mechanism behind KSS-LoRA - and the math that makes it provably stable at any precision. Gradient Underflow in Reduced Precision Training: A Deep Dive https://koscak.ai/blog/gradient-underflow-reduced-precision/ https://koscak.ai/blog/gradient-underflow-reduced-precision/ Why gradients vanish at FP8 and the cascading effect on model quality. The bit-level explanation. Overfitting in LLM Fine-Tuning: What the Loss Curve Isn't Telling You https://koscak.ai/blog/overfitting-llm-fine-tuning-loss-curve/ https://koscak.ai/blog/overfitting-llm-fine-tuning-loss-curve/ Your validation loss looks fine. Your model is ruined. The hidden overfitting problem with exact numbers. NVIDIA Hopper vs Blackwell: Training Performance Breakdown https://koscak.ai/blog/hopper-vs-blackwell-training-performance/ https://koscak.ai/blog/hopper-vs-blackwell-training-performance/ H200 vs B300 - architecture differences that matter for fine-tuning, with real training metrics. NVIDIA B300 Blackwell Ultra: First Fine-Tuning Results and Scaling Curve https://koscak.ai/blog/nvidia-b300-blackwell-ultra-fine-tuning-results/ https://koscak.ai/blog/nvidia-b300-blackwell-ultra-fine-tuning-results/ First published fine-tuning results on B300 - 7-model scaling curve from 1B to 72B parameters. Llama 3.1-8B Fine-Tuning: Best Practices in 2026 https://koscak.ai/blog/llama-3-fine-tuning-best-practices-2026/ https://koscak.ai/blog/llama-3-fine-tuning-best-practices-2026/ Everything from 88+ training runs on A100 and H200. The configs that matter and the ones that don't. The Hidden Cost of FP8 in Production AI Factories https://koscak.ai/blog/fp8-production-ai-cost-quality/ https://koscak.ai/blog/fp8-production-ai-cost-quality/ Every hyperscaler running fine-tuning at FP8 scale is quietly losing quality. Here's the cost math. Qwen2.5-7B vs Llama 3.1-8B: Fine-Tuning Comparison https://koscak.ai/blog/qwen2-vs-llama-fine-tuning-comparison/ https://koscak.ai/blog/qwen2-vs-llama-fine-tuning-comparison/ 46 identical experiments on both models. The numbers on which base model wins and when. HuggingFace PEFT: What the Docs Don't Tell You About LoRA https://koscak.ai/blog/huggingface-peft-lora-hidden-gotchas/ https://koscak.ai/blog/huggingface-peft-lora-hidden-gotchas/ Default PEFT config works on A100. On H200 FP8 it silently destroys your model. Here's exactly what to change. LoRA Rank Sensitivity: How r Affects Your Training Outcome https://koscak.ai/blog/lora-rank-sensitivity-guide/ https://koscak.ai/blog/lora-rank-sensitivity-guide/ 8 rank values tested across 46 runs. The optimal value and why bigger isn't better. TruthfulQA as a Fine-Tuning Benchmark: Why It Catches What Other Metrics Miss https://koscak.ai/blog/truthfulqa-fine-tuning-benchmark/ https://koscak.ai/blog/truthfulqa-fine-tuning-benchmark/ Standard loss metrics look healthy while the model degrades. TruthfulQA catches it. Here's the data. Antifragility in Machine Learning: Taleb's Framework Applied to Neural Networks https://koscak.ai/blog/antifragility-machine-learning-neural-networks/ https://koscak.ai/blog/antifragility-machine-learning-neural-networks/ Taleb's antifragility concept applied to training - and experimental proof it's achievable. Numerical Stability in Low-Precision LLM Training: A Practical Guide https://koscak.ai/blog/numerical-stability-low-precision-llm-training/ https://koscak.ai/blog/numerical-stability-low-precision-llm-training/ Why low-precision training breaks more than the papers admit - and the constraints that fix it. Unsloth Fine-Tuning Guide: Speed Up Training Without Losing Quality https://koscak.ai/blog/unsloth-fine-tuning-guide-2026/ https://koscak.ai/blog/unsloth-fine-tuning-guide-2026/ 2-3× faster fine-tuning - and the settings that actually matter vs ones you can ignore. RunPod vs Lambda Labs vs Vast.ai: GPU Cloud for AI Training in 2026 https://koscak.ai/blog/runpod-vs-lambda-vs-vast-gpu-cloud-2026/ https://koscak.ai/blog/runpod-vs-lambda-vs-vast-gpu-cloud-2026/ Real costs and real reliability after 88+ training runs across three providers.