We built Neptune Scale to let you monitor such training and debug any issues quickly. Now available in beta: https://t.co/7lWqtFSm7g Coming soon for everyone.

Track down gradient explosions to the exact layer and training step. — Watch the full product demo: https://t.co/xGvTQ78QwB

When you’re tracking experiments, benchmarks matter A LOT. With Neptune, you can now overlay baselines, thresholds, or success criteria directly on your charts. No more mental math or custom expression workarounds. Just clear, reliable context right in the plots.

[Editor's Pick] Learnings From Teams Training Large-Scale Models: Challenges and Solutions For Monitoring at Hyperscale Author: Siddhant Sadangi Reading time: 5 min — Full article: https://t.co/W6VoKxsJmf

Mixture-of-Experts can save compute or burn it. At Qualcomm, research intern Maciej Pióro ran a grid of #MOE experiments that spiraled into 50,000 A100 hours of re-training. - Why: The load-balancing loss was averaged across layers, so deeper experts ignored it. Tokens

The Neptune chart legend has come a long way to become a real tool for inspecting and navigating run results: → Started as a simple floating label → Now: resizable, draggable, pinnable → Add params, tags, or any other attributes → Search through it easily → Use it to

[Editor's Pick] Hyperparameter Optimization For LLMs: Advanced Strategies Authors: Gabriel Souto Augusto Dutra, Kilian Kluge Reading time: 14 min — Full article: https://t.co/S8Im20OCGa

Spot anomalies at 600K (or any other) steps without zooming in or digging through logs. — Watch the full product demo: https://t.co/xGvTQ78QwB

Not all layers learn at the same rate. Some layers in deep models "go silent" during training—gradients shrinking to near-zero. Others behave erratically, with values jumping unpredictably. Both cases harm convergence, and neither is obvious from loss curves alone. That’s why,

Now in Neptune: Shared zoom. Shared cursor. Shared highlights. These should help you stay oriented when you’re debugging and the metrics volume gets real.

[Editor's Pick] Open LLMs are Necessary For Current Private Adaptations and Outperform Their Closed Alternatives Author: Olatunji Iyiola Emmanuel Reading time: 5 min — Full article: https://t.co/Qez4mcpmWO

Scatter plots make it easy to spot trends. In this example, a look at BLEU scores and a config parameter shows that the best models used the lowest training floor. — Discover more ways to analyze your experiments with Neptune: https://t.co/xGvTQ78QwB

Every researcher has a story that starts with: “We thought everything was fine… until it wasn't.” At ICML 2025, we turned those moments into a series, TrainFM, where top researchers share their biggest training failures and how they fixed them. It’s raw. It’s messy. And

When you’re tracking experiments, benchmarks matter A LOT. With Neptune, you can now overlay baselines, thresholds, or success criteria directly on your charts. No more mental math or custom expression workarounds. Just clear, reliable context right in the plots.

Monitor → Spot spikes → Debug → Fork In this short demo, you can see how to do the forking with Neptune. Update training config, and continue from a stable checkpoint. The forked run inherits metrics up to the split, giving you a complete view of both runs side by side. —

[New on our blog] How to Optimize LLM Inference Author: Alek Pikl Reading time: 12 min — Full article: https://t.co/CiPTdvUmmv

A slow tracker kills research momentum. It made our day to hear that Alex Immer, a Senior Research Scientist at Bioptimus, was “very pleasantly surprised” by Neptune’s performance. Lightning-fast UI, real-time logging, no overhead, even at large scale. — Watch how Bioptimus

Metric blindness can cost you hundreds of GPU days. That’s the painful lesson Saaketh Narayan from Meta's Llama pre-training team shared with us. During FP8 training, unnoticed gNorm anomalies led to catastrophic loss spikes, wasting ~512 GPU days. The fix? Architectural

We’ve made a lot of small improvements to @neptune_ai’s chart legend over time. Together, they’ve added up to something that’s quietly powerful. → Floating or attached to the bottom of the chart → Resizable, draggable, pinnable → Search inside the legend → Add custom

“Usually, the first idea doesn't check out. That's just the nature of #AIresearch.” For Derek Li, Senior Researcher at Noah's Ark Lab, that first idea was training all multitask #reinforcementlearning objectives together. What went wrong: Runs underperformed baselines, with

Before: Shaded regions in @neptune_ai = min–max range. Now: You decide. Custom Error Bands let you bring your own error metrics like standard error, confidence intervals, or any bounds you care about.