Insights, tips, and product updates

ImpactBench scored ten models on 809 expert-sourced wellbeing metrics, then had experts audit its LLM judge. Here's what the 83.3% agreement figure hides.

Decision models give LLM judges calibrated probabilities, but model calibration is only ever measured against someone's labels. See whose labels the new models use, and why yours matter more.

SFT vs RLHF is usually framed as cost against capability, but it is also a staffing decision. This guide shows which stage needs experts who can do the task, which needs experts who can judge it, and how to check the judges.

Your labeling platform is already an AI model evaluation workflow with a different input. Here is the six-step loop, and the evidence for why teams break it.

Model distillation trains a small, cheap model on the outputs of a larger one, and the student inherits the teacher's errors along with its skills. Learn how knowledge distillation works and where to check the teacher's outputs before they become training data.

Two 2026 studies show that LLM-as-a-judge setups can pass every consistency check and still disagree with the experts they are meant to stand in for. Learn which agreement metrics expose the gap, and why the annotator pool behind your "ground truth" matters as much as the judge.

RLVR trains large language models on rewards a program can check, such as a correct maths answer or a passing unit test. This guide explains how reinforcement learning with verifiable rewards works, how models learn to game the verifier, and how to audit one before you trust it.

GRPO is the reinforcement learning algorithm behind DeepSeek-R1, Qwen3 and DeepSeek-V3.2. This guide explains how group relative policy optimization scores multiple responses against each other, and why your prompts and your reward model decide whether it learns.

RLHF is now one stage in a three-stage post-training pipeline, and each stage asks people for a different kind of judgment. This guide maps SFT, preference optimisation and verifiable rewards by the human work they need, and by how each one fails when that work is weak.

Xiaomi's MiMo-V2.6 report shows that 12.7% of a $2.6 million RL post-training run went to grading, with an LLM-as-a-judge at almost every layer. This Data Story maps who judges what in that pipeline and the one measurement the report leaves out.

An AI benchmark audit moved one model's SWE-Bench Pro score by 19 points without touching the model. This is what the re-annotation changed, and what the eighteen months before it revealed.

A stage-by-stage record of Granite 4.2 training data, from pretraining phases to reinforcement learning. Every figure carries the IBM document it came from.

Golden datasets for testing, fine-tuning, and evaluation are three different artifacts. Each has its own quality bar, sourcing method, and failure modes. This guide covers what makes each type effective, where teams go wrong, and how human-in-the-loop annotation infrastructure turns quality principles into measurable, governed datasets.

Fine-tuning audio models — ASR, diarization, speech-to-text — is bounded by your reference transcripts. What to annotate, in what order, and how to avoid circular training data.

Audio annotation turns recordings into verified ground truth for speech AI. This guide covers the types, process, costs, quality control and EU AI Act rules.

Compare the best computer vision annotation tools that run on infrastructure you control, evaluated on deployment architecture, video and 3D coverage, and whether label quality is built into the workflow or measured after it. This guide covers seven platforms, names the three that no longer qualify, and shows where the line between self-hosted inference and self-hosted annotation actually falls.

Published word error rates come from read audiobooks, close-talk meeting headsets and telephone corpora three decades old. This guide takes apart the distance between those numbers and the transcript of an actual customer call: what WER counts, what it forgives before counting begins, and what it cannot see.

Speaker diarization looks solved on benchmark corpora and clearly is not on real recorded calls. This guide explains what diarization error rate actually measures, why the figures aren't comparable across sources, and which failure modes benchmarks exclude.

Gemma 4 added 4B parameters over Gemma 3 and quadrupled its AIME score. The technical report credits data composition for the gain, then describes its training data in two sentences.

Kimi K3 ranked third on the AI Intelligence Index while its hallucination rate hit 51%. Here is what that paradox reveals about how the industry evaluates models.