hrm-text-rsi-sampled-20260923

# HRM-Text sampled pretraining data A fixed four-epoch dataset for HRM-Text PrefixLM pretraining, with a shared compact token array and the tokenizer needed to read it. ## Contents - `sampled/tokens.npy`: a memory-mappable `uint16` token array. - `sampled/epoch_0` through `sampled/epoch_3`: instruction and response start offsets and lengths, stored as NumPy arrays. - `sampled/metadata.json`: the context limit, epoch length and tokenizer settings. - `tokenizer/`: the frozen 65,536-token BPE tokenizer. Each epoch contains 58,877,637 sampled rows. The shared token array contains 42,595,442,478 tokens. The model context is 4096 tokens; the data context bound is 4097 to include the autoregressive shift. ## Download The complete download is approximately 93 GB. ```python from huggingface_hub import snapshot_download snapshot_download( repo_id="sapientinc/hrm-text-rsi-sampled-20260923", repo_type="dataset", local_dir="hrm-text-data", ) ``` Use an authorized Hugging Face login if authentication is required. ## Use Use `epoch_0` for one epoch, the first two index sets for two epochs, and all four index sets for four epochs. Reuse these fixed index sets without resampling. Later epochs can include additional documents. Load the arrays with the HRM-Text PrefixLM data loader and the included tokenizer. The runtime metadata uses `/datasets/hrm-text/tokenizer` as its container tokenizer location; applications using a different layout can configure the tokenizer location locally. This repository contains training data and a tokenizer. Pretrained model weights and evaluation artifacts are not included.