3LC

Training or finetuning Machine Learning Models? Record Per-Sample Per-Epoch Metrics and Visualize and Improve your data! We're seeking valuable feedback from the Data-Science and Machine Learning Community, helping to shape what we believe is a unique product. 3LC integrates with PyTorch and your existing Training Scripts. We believe we’re unique in that we record in detail what goes on during your training, for each sample for each epoch. And then you can visualize it with high-end visualization, locally, during or right after your training run. For example, visualize the distance training samples were moving in embedding space throughout different epochs, how often did they swap clusters (noisy and hard samples!)? Or visualize how each prompt over every epoch in RLHF behaves (Finetuning a Hugging Face Model with a Prompt dataset), cross-plotting loss, chosen_award and rejected_award. You can lasso in on samples and see the actual underlying data and understand what to do to improve the model. One thing we believe is useful, is that you can interactively edit the dataset samples in the UI (single or in batch), and thus change your data before retraining again. These are just a few examples of what you can do in the UI and with 3LC. We want to ensure that these tools are accessible to everyone, that's why 3LC is free for non-commercial use! As we roll out the beta version of 3LC, we invite 100 of you at first, to explore its unique capabilities, with minimal intrusion into your existing workflows(we hope!). Please remember we’re continuously developing and improving 3LC, it's far from finished but we want your feedback to help make something powerful. (There will be bugs... and the visualization might need a discrete graphics card now, dependent on your hardware.) pip install 3lc https://docs.3lc.ai/ Looking forward to your feedback!