This public pilot contains packed uint16 token IDs for pretraining the Tiny Base 300M, 2,048-context language model. It contains 50,000,000 training, 1,000,000 validation, and 1,000,000 test tokens. It does not redistribute raw documents. train.bin, validation.bin, and test.bin are little-endian uint16 streams. Use tokenizer.json with the Hugging Face tokenizers library. manifest.json records exact token counts, split policy, source revisions, mixture, and SHA-256 checksums; settings.json records preparation settings. The deterministic split is SHA-256 of normalized text modulo 1,000: 0–9 test, 10–19 validation, and 20–999 training. The mixture is 70% FineWeb-Edu, 25% English Wikipedia, and…