Better models start with better examples.
Bring your documents, conversations, and Hugging Face datasets into one review workflow. Turn the examples you trust into a version you can train and evaluate.
Import existing examples, or generate them from source documents with a recipe.
Interactive illustration · sample data
From source to training set. Follow an example through import, review, and a reproducible dataset version.
Know what your model is learning.
Keep source material, review decisions, and training data connected. A larger dataset only helps when you can trust what is in it.
Bring the data you already have
Upload files, paste content, or import a Hugging Face dataset. Preview rows and map your columns to conversations or chosen/rejected preference pairs.
Explore the featureGenerate examples with evidence
Use recipes for factual Q&A, troubleshooting, multi-turn conversations, or structured JSON. Generated examples carry quotes from their source passages into review.
Explore the featureReview with context
Edit answers, inspect quality signals, find similar prompts, and approve or reject examples. Review individual conversations or work through a filtered selection.
Explore the featureFreeze a reproducible version
Create a snapshot of approved examples. Related sources, prompts, and images stay in the same split to help prevent training and test leakage.
Explore the featureCheck the data before you rent the GPU.
Dataset readiness checks use the model’s tokenizer to surface empty answers, invalid conversations, and examples that exceed the training window. The check starts no GPU and calls no paid model.
- Source-aware
- Trace generated answers back to the passages they came from.
- Test data stays held out
- Preserve split assignments across later dataset versions.
- Your data, portable
- Download training, validation, and test splits as JSONL.
Build with Tensorant