Qwen3-8B faithful-HQMQ vs E8 Pareto comparison

[jagmarques]

Summary

Extends the faithful-HQMQ Mistral kernel to Hurwitz's own evaluation model (Qwen3-8B). Measures the PPL-vs-bpe Pareto frontier for HQMQ (s96_r6 ~5 bpe, s96_r4 ~3.79 bpe) and E8 (K4V2, K3V2) under identical harness conditions.

• HQMQ algorithm unchanged from spec (24*S product code, Med3x C=3, no rotation, per arXiv:2605.27646)
• Model: Qwen/Qwen3-8B with NF4 weights (fits T4 x2); all configs share the same weight baseline
• Paired PPL harness: n=60 WikiText-103 segments, prefix=1024, cont=512 tokens
• Reports measured bpe per config (not hardcoded nominals); Pareto comparison honest on both axes

Status

GPU result pending. No PPL claim is made here. Draft until jagmardrop/nq-hqmq-qwen3 completes.

Proof of work

VERIFY-WITH: python3.11 experiments/kaggle/nq_hqmq_qwen3/smoke_test.py

TEST (a): codebook size = 24*S ... PASS (S=24,48,96,192 all have 24*S unit-norm codewords)
TEST (b): round-trip bounded error ... orig_norm=5.8334, err=0.4811, ratio=0.0825 PASS
TEST (c): Med3x outlier detection ... PASS (1 block(s) flagged; planted at idx 31 norm=10.50 > 3*1.00)
TEST (d): HQMQ bpe s96_r4 ~3.79 ... base_bpe_r4=3.7925 corrected(p=0.02)=4.2866 PASS
TEST (e): HQMQ bpe s96_r6 (base formula, no outlier correction) ... base=4.2925 with_flag=4.5425 corrected(p=0.07)=5.3620 PASS
TEST (f): E8 bpe formula K4V2 and K3V2 ... K4V2=2.0000 K3V2=1.9375 PASS (K4V2=2.000 > K3V2=1.938)
All smoke tests PASSED.
exit code: 0

Kaggle push: jagmardrop/nq-hqmq-qwen3 version 1 (RUNNING at push time)
https://www.kaggle.com/code/jagmardrop/nq-hqmq-qwen3

Diff stat: 3 files changed, 806 insertions(+)