Recipes minicpm-v-4.6 (dense shares, vocab_sweep 11) and minicpm-v-4.6-moe (MoE shares), rendered through openbmb/MiniCPM-V-4.6@36f34a66's own template (auto_fmt), its vocabulary sweep (248094 rows). Pool documents sharing a 13-gram with an eval file were quarantined (purge_overlap) before the final build. For the MiniCPM-V 4.7 release if its template and added tokens fingerprint as 4.6's.
Restructure into a pool + per-model builds; add Muse-Glimmer-30B build
The flat corpus could not survive a change of model: DeepSeek-V4 chat markup was
baked into the agentic and reasoning slices and the vocabulary sweep was built
for DeepSeek's 129,280 embedding rows. Against Muse-Glimmer-30B it covered
47.56% of 202,048 rows; rebuilt from the pool it covers 97.22%.
pool/ raw units with provenance and licence, dialogue stored as messages.
recipes/ what a build is made of. builds/muse-glimmer-30b 4.90M tokens plus a
751k-token long-context file. eval/ three measurement corpora, 0 shared
13-grams with any build and with each other; the old eval_neutral.txt is
renamed to eval/code/eval_code.txt byte-identically because measured, it is a
code corpus. tools/ the pipeline.
llama-imatrix needs --parse-special or the markup tokenizes as literal
punctuation. llama-tokenize aborts on some plain-ASCII input (invalid codepoint
in src/unicode.cpp) which the previous calib_train.txt would have triggered.
