Weights & Biases

“I use WandB in my job multiple hours per day. It's the most feature-complete thing for this application out there, but its performance is soooooo gratingly bad.”

“With how popular LLMs are, its mindboggling how wandb lacks a feature for continuously logging text generation samples during training. I'm aware they suggest using tables for logging text data, but from my understanding they are not designed to be updated. So each time you want to log new text samples, you have to create a new table. And for some reason the feature request of this issue is closed due to stale , and the community is doing gymnastics to overcome this problem.”

“Wandb only logs what you tell it to log. The data hosted on their server is opt-in, meaning you explicitly log what you want when you use their API. If you don't trust any third party with your process, the local setup seems pretty straightforward (docker container), info at https://docs.wandb.ai/guides/hosting/basic-setup. I did use some other experiment tracking software and found wandb to be easier and full of useful features. Granted I didn't try MLFlow, but wandb had almost everything I nee”

“W&B is fantastic and comet.ml is even better if you're trying to do ML experiments at scale. like kubernetes, docker, a linter, git, pull requests, pre-commit hooks: you're not going to see the value at small scale. If you're just one person you probably do think "this sucks" and thats ok. then you try and have 100 engineers working on the same codebase and you go OH. YEAH. LETS HAVE A LINTER. Langchain just sucks. Its truly awful. And its also not the type of tool like linters or experiment tra”

“What stands out most about Weights & Biases is how seamlessly it tracks ML experiments through a clean, intuitive UI/UX. It integrates effortlessly with frameworks like PyTorch and Hugging Face, which makes real-time performance monitoring and fast, reliable data logging feel almost automatic. Onboarding is quick and well supported by strong documentation, so it’s easy to get up and running without friction. The AI intelligence features also help keep artifact management organized and make hyper”

“I keep bumping into these tools (weights & biases, langfuse, langchain) and honestly I’m not sure if it’s just me but the UX feels… bad? Like either bloated, too many steps before you get value, or just generally annoying to learn. Curious if other engineers feel the same or if I’m just being lazy here: • do you actually like using them day to day? • if you ditched them, what was the dealbreaker? • what’s missing in these tools that would make you actually want to use them? • does it feel like t”

“Weights & Biases makes MLOps and experiment tracking feel effortless, with only minimal code required to get started (wandb.init()). Its real-time, interactive dashboards deliver immediate and detailed visualizations of training loss curves, the impact of hyperparameters, and system resource utilization (GPU/CPU). On top of that, robust artifact versioning and seamless integrations with frameworks like PyTorch, Hugging Face, and TensorFlow make model reproducibility and team-wide collaboration e”

“I primarily use Weights & Biases to track and compare machine learning experiments, monitor training metrics in real time, and manage model versions. I really like how it solves the challenge of keeping experiments organized and reproducible, with everything logged automatically. What I like most about Weights & Biases is how effortless it makes experiment tracking and visualization. The interactive dashboards, real-time training metrics, hyperparameter comparison tools, and artifact management ”