Hey Product Hunt! 👋

I'm Abdullah, the creator of Gradian.

Over the past few months, I've been spending a lot of time fine tuning LLMs, and I kept hitting the same frustrating problem. The model would get worse, but I had no idea why.

Nothing crashed. The loss curve looked normal. Training finished successfully. But suddenly the model would stop mid sentence, lose capabilities it had before, or start giving confidently wrong answers.

Debugging meant changing one thing at a time, retraining, and then waiting hours to see if I had guessed correctly. Most of the time I hadn't.

So I started building the tool I wished I had.

The first question I wanted to answer was simple. Which training examples actually caused a capability to improve or get worse?

That led me into gradient based data attribution. Along the way I realized many of the biggest issues had nothing to do with the data itself. They came from subtle training mistakes like truncated completions, missing EOS tokens, loss being computed on prompts, train and eval contamination, or hyperparameters copied from a completely different training recipe.

The biggest surprise came while building the attribution method itself. My first version consistently ranked intentionally poisoned examples as some of the most helpful examples in the dataset. It turned out they were teaching the model the output format perfectly, even while teaching the wrong facts. That forced me to rethink the approach, and eventually led to subtracting the gradient of the model's actual prediction. The rankings immediately started matching reality.

Gradian is the result. It's an open source tool that helps explain why a fine tune changed, instead of leaving you guessing.

I'd love to hear your stories. If you've ever had a fine tune mysteriously get worse, what ended up being the cause?

Thanks so much for checking out Gradian! 🚀