allocateFine-tuning dataset validator
Paste or drop your JSONL training file. Format checks, per-line errors, token counts, and a cost estimate; nothing ever leaves your browser.
How it works
Common questions
Is my training data uploaded anywhere?
No. The validator runs entirely in your browser; your dataset never touches a server. That matters when the file contains customer records.
What format does it expect?
The standard chat fine-tuning format: one JSON object per line with a messages array, each message carrying a role (system, user, assistant, or tool) and string content. This is the format OpenAI-compatible fine-tuning APIs and open-weight trainers accept.
How many examples do I need to fine-tune?
Quality beats volume. A few hundred excellent examples move behavior; a few thousand resolved real cases usually beat any amount of synthetic data. The validator's token count tells you what an epoch will cost.
How accurate is the token count?
It uses the standard four-characters-per-token approximation, typically within 10 to 15% of exact tokenizer counts across models, which is enough for cost planning.
What do I do after validating?
Estimate the training run with the fine-tuning cost calculator, then pick an open-weight base from the model library. On Allocate the trained weights stay inside your boundary and belong to you.
More free tools
Dataset clean? Allocate fine-tunes open-weight models on it inside your own boundary, and the weights belong to you.