climbmix_phrase_v1

ClimbMix (karpathy/climbmix-400b-shuffle, the same corpus split and val text as ChrisMcCormick/climbmix_unigram16k_14_170) re-tokenized with a phrase vocabulary of 1,483,965 ids built on top of priml's 16,384-piece byte-level Unigram: the 16K pieces keep their ids, and 1,467,581 phrase entries follow, each a sequence of 2-8 pieces. Built by rounds of batched pair merges over the 16K id stream of train shards 1-10 (1.0B tokens): frequency order, a PMI floor of 1 bit, min count 32, 150K merges per round, at most 8 pieces per entry, then entries used fewer than 32 times fall back to their parents (56,253 of them). Held out, the val shard goes from 57.0M to 26.6M tokens: 9.50 bytes per token against the 16K tokenizer's 4.43 (2.14x); 73% of the entries span a word boundary.

• climbmix/train_0000NN.bin, climbmix/val_000000.bin: decoder-rtx's .bin layout with a 256-int32 header (magic 20240520, version 2, num_tokens) and int32 tokens (version 1 is the uint16 layout). Each document opens with BOS (id 16,368). Train shards 1-10 are the build corpus; 11-13 were rewritten through the same vocabulary without being part of it.
• vocab/v1-pmi1-32.npz: entry_pieces / entry_offsets (each phrase id's 16K pieces, in dense id order from 16,384), dense_nbytes (bytes per id, the decoded-token convention), dense_npieces, and everything the builder's apply-only mode needs to rewrite more shards (parents, pruned, remap, the per-round merge tables). vocab/v1-pmi1-32.json: the build summary. Decode an id by decoding its pieces with the 16K tokenizer.json.

Builder and logs: agent-ops-stacks/encoder-rtx/2026-11-08_0739am_phrase-vocab-sampled-head/ (build_phrase_vocab-01-gpu.py, box_v1.sh, box_rewrite.sh).