For Aleph Alpha's Kolibri-1 (78B MoE, trained on 23.9% German).
- pool/german/ (tools/pool_german.py): Wikipedia DE, OpenAssistant oasst2 DE, ultrachat_de, germanrag,
templated German tool-calling dialogues, long German articles for calib_longctx. Licences in every record.
- pool/vocab_sweep/synthetic/kolibri-1-moe.jsonl: 127,148 of 127,166 sweepable ids.
- recipes/kolibri-1-moe.yaml and builds/kolibri-1-moe/: 5,598,775 tokens, 4,773 documents, German 21.4%,
reasoning_effort rendered through the model's template (tools/build.py + tools/auto_fmt.py); calib_longctx
771,843 tokens in 29 documents. The qwen3.8-flash-next-moe recipe rebuilds byte-identical with these tools.
- eval/neutral-de (348k tokens), eval/longctx (349k, eight chunks at 32k or needles to 128k),
eval/agentic-kolibri (354k: eval/agentic's generator, scripted and grounded, in Kolibri-1's template;
tools/build_eval_auto.py, repository commits in grounding.json).
- builds/kolibri-1-moe/crosscheck.json: no 13-gram…
feat(builds): qwen3.8-flash-next-moe, the MoE build for qwen4exp
- recipes/qwen3.8-flash-next-moe.yaml: MoE shares 22/16/12/18/10/12/6/4, llama.cpp 957538960
- builds/qwen3.8-flash-next-moe: 4,846,205 tokens, 2,755 docs; coverage 99.974% of reachable ids; 0 shared 13-grams with eval neutral/code/code_ext
- pool/vocab_sweep/synthetic/qwen3.8-flash-next-moe.jsonl: sweep for the Flash-Next tokenizer (the August qwen3.8-flash-next sweep never reached the pool)
- pool: 59 Wikipedia articles and 1 code file that eval_neutral / eval_code_ext share text with moved to _quarantine/eval-residue.jsonl
- tools/purge_overlap.py: --whole-pool; append to the residue shard instead of overwriting it (each run dropped what earlier runs had quarantined)
builds/qwen3.8-flash-next (August, dense shares, identical to qwen3.8-27b) is left as it was.
:hammer: feat(tools): Add abliteration script
worthant
814d662
refactor(tools): move clone.sh into tools and drop the hardcoded path
worthant
3ab87b6
The raw directory came from a literal /home/boris path, so the pool couldonly be harvested on one machine. It now takes the directory as anargument or from CALIB_RAW.Repositories stay unpinned on purpose: builds are reproducible from thecommitted pool, not from these clones, and fresher upstream code isbetter material. What was missing was the record, so clone-manifest.tsvnow carries the commit, its date and the licence for each repository.
feat(tools): ship the auto renderer inside the dataset
worthant
0eba5ab
build.py knew only glimmer and nemotron, so a recipe withchat.format: auto stopped with 'unknown chat.format'. The file livedonly in the quantizer repo and foundry patched build.py with sed onevery box, which meant a clean clone of this dataset could not rebuildthe Qwen corpus at all.auto_fmt.py now lives in tools/ and is registered in FORMATTERS.install_auto_fmt in foundry.sh is no longer needed.
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.