int32 (N,512) Qwen2.5 token ids (vocab 151936), our own derivation:
poolv3_lite 9,838 x 512 (wiki 49.6 / code 20.6 / chat 9.4 / math 11.9 / uber 8.5)
poolv3b 39,130 x 512 intermediate pool
poolv3b65 53,702 x 512 the general pool (wiki-anchored)
code_mix_15b 15,993 x 512 wiki+ubergarm+DPQ high-doubt code/math/chat
general_hs_15b 56,202 x 512 poolv3b65 + 2,500 HellaSwag-train blocks
mix_wiki_uber 1,670 x 512 wiki+ubergarm 50/50 (M1 arm)
onpolicy_r1 1,500 x 512 student's own generations (on-policy distillation)
Text is NOT rehosted; the per-source license map is LICENSE-DATA.md and the
composition/provenance is PROVENANCE.md.
Published so the benchmark and calibration claims are checkable: MBPP item ids
(all-427, the 0.5B and 1.5B F16-solved subsets), the byte-identical Apache-2.0
sanitized-mbpp.json, the harness that produces the quoted accuracies, the
quantization stage drivers, per-file provenance with sha256, and the fetch script
for the upstream text we do not rehost. Calibration mixes follow in a second
commit.
