FP8 gradient underflow, LoRA overfitting, antifragile training on NVIDIA Blackwell. We ran the experiments so you don't have to guess.
Verify or abstain, running live: how a zero-LLM engine answers only what it can prove, and keeps its lie counter at zero.
One station, nine years of data, one Raspberry Pi - measured against the operational global models on held-out 2026 data.
Scanner output is not an assessment. Chained-finding analysis under exact-or-abstain reporting - starting with a free scan.
Standard LoRA at FP8 loses 68% quality without any warning signs. Here's the root cause and the fix.
50% noise injection reduces the overfitting gap by 94.7%. Most methods break under noise. This one uses it as fuel.
Complete guide to fine-tuning on B300 Blackwell Ultra with FP4 precision. Setup, config, and results.
The Bernoulli sparsity mechanism behind KSS-LoRA - and the math that makes it provably stable at any precision.
Why gradients vanish at FP8 and the cascading effect on model quality. The bit-level explanation.
Your validation loss looks fine. Your model is ruined. The hidden overfitting problem with exact numbers.
H200 vs B300 - architecture differences that matter for fine-tuning, with real training metrics.
First published fine-tuning results on B300 - 7-model scaling curve from 1B to 72B parameters.
Everything from 88+ training runs on A100 and H200. The configs that matter and the ones that don't.
Every hyperscaler running fine-tuning at FP8 scale is quietly losing quality. Here's the cost math.
46 identical experiments on both models. The numbers on which base model wins and when.
Default PEFT config works on A100. On H200 FP8 it silently destroys your model. Here's exactly what to change.
8 rank values tested across 46 runs. The optimal value and why bigger isn't better.
Standard loss metrics look healthy while the model degrades. TruthfulQA catches it. Here's the data.
Taleb's antifragility concept applied to training - and experimental proof it's achievable.
Why low-precision training breaks more than the papers admit - and the constraints that fix it.
2-3× faster fine-tuning - and the settings that actually matter vs ones you can ignore.
Real costs and real reliability after 88+ training runs across three providers.
