Record: Trinity v7+skip — val_bpb 0.22311 (3-seed mean, NEW #1)

[deborahnelson8788726]

🏆🏆 MASSIVE UPDATE: val_bpb 0.22311 (3-seed mean)

N-gram Entropy Skip = -33.5% BPB improvement

Seed	val_bpb
42	0.22509
314	0.22253
999	0.22172
Mean	0.22311

Key insight: when n-gram is confident (p>0.8) AND neural model uncertain (entropy>1.5), skip blending and use pure n-gram. Avoids diluting near-perfect predictions.

• vs SOTA (1.081): -79.4%
• vs PR Record: Per-Sample SLOT + N-gram Order-22 + TTT + LR=0.432 — val_bpb 0.39642 (3-seed mean) #1430 (0.396): -43.7%

One line of code, biggest single improvement in the project.

[MatoTeziTanka]

Community Review — Record: Trinity SLOT v3 + Pre-Quant TTT — val_bpb 0.65802 (3-seed mean)

BPB: 0.65802 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA 7141d2af5882, file records/track_10min_16mb/2026-04-02_Trinity_Hybrid_Ternary_GPTQ_XSA/train_gpt.py):

The TTT path at line 1145 implements the score-first-per-chunk pattern: each chunk is scored under torch.no_grad() / inference_mode() before the base_model.train() + SGD adaptation runs on that same chunk, with an is_last_chunk guard so the final chunk gets no adaptation pass. This is the structural shape the legal frontier uses (PRs #1416 erichroepke, #1423 aryanbhosale).

Per Issue #402 and Issue #677, TTT is legal when each token is scored before the adapter updates on it, and that's what the code does here — chunk ci is scored under weights adapted only on chunks 0..ci-1. No prequant_ttt_adapt_adamw(val_tokens, ...) multi-epoch fine-tune, no scored-region SLOT, no target-in-key n-gram cache.

CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.04s, dim=512, layers=11, vocab=1024, code=126681 B, SMOKE_TEST_PASS

Verdict: LOOKS CLEAN.

Recommendation to @cocohearts @valerio-oai @0hq @yuzhougu-oai @notapplica: MERGE pending standard checks (3-seed validation, 16MB artifact cap, 10-min wallclock on 8×H100 SXM). The compliance picture matches the legal reference frontier and no flags were raised by the classification pass.

Auto-classification caveat: this review was drafted by the AST-based classifier against a template derived from manually-reviewed cluster PRs (#1420, #1450, #1487, #1541, #1529, #1533, #1518). If I've misread a subtlety in your eval path — e.g., multi-epoch TTT that I mistook for single-pass, or a target-in-key lookup I missed in a helper function — please flag it and I'll re-run the audit manually.

---

Reviewed by @MatoTeziTanka — The Agora. CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.04s, dim=512, layers=11, vocab=1024, code=126681 B, SMOKE_TEST_PASS. Classification via deterministic AST-based classify_prs.py (pattern bank derived from ~65 manually-reviewed PRs earlier in the 2026-04-11 sweep). This review was auto-drafted from a template and spot-checked before posting — if the template misread your code, please call it out so I can iterate the classifier.

[deborahnelson8788726]

🏆 NEW RECORD: val_bpb 0.37112

Trinity v6 = N-gram Order-22 + Per-Sample SLOT + Pre-Quant TTT

Single seed 42 on 4xH100 SXM:

• val_bpb: 0.37112 (beats PR Record: Per-Sample SLOT + N-gram Order-22 + TTT + LR=0.432 — val_bpb 0.39642 (3-seed mean) #1430's 0.39642 by 0.02530)
• Beats official SOTA (1.0810) by 65.7%
• Training: 2762 steps, 600s
• Total eval: ~1500s (TTT 703s + SLOT+N-gram 785s)

Key innovation: GPU-vectorized Backoff N-gram Order-22 mixer with entropy-adaptive blending on top of Per-Sample SLOT (LR=0.432, beta1=0.6, beta2=0.5).

Latest commit: a18c7ef
Trinity framework: https://github.com/gHashTag/trinity

[deborahnelson8788726]

Closing this PR after a careful re-read of Issue #677 and #1017.

The hash-based N-gram + entropy skip approach used here violates Condition 2 (full-vocab normalized distribution): BackoffNgramMixer.score() computes ngram_p only for the target token via self.uni_counts[targets], never building a full V-dim distribution that sums to 1. The reported jump from 0.336 → 0.223 BPB after enabling the entropy skip is largely an artifact of hash collisions in the bucketed n-gram cache, not real predictive gain — consistent with @Eppie's analysis in #677 and the empirical sweep in PR #886 showing the apparent improvement collapses to ~0.0002 BPB when collision-free buckets are used.

The Per-Sample SLOT cluster (PR #1329, #1240, #1336) is also under hold/closure and our SLOT pattern matches the flagged structure.

Closing self-initiated rather than wait for the inevitable. The experimental code remains in git history for reproducibility / as a cautionary example.

Cleaner Trinity submissions:

• Record: Trinity SLOT v3 + Pre-Quant TTT — val_bpb 0.65802 (3-seed mean) #1722 — Trinity SLOT v3 + Pre-Quant TTT, val_bpb 0.65802 (3-seed mean), community-reviewed LOOKS CLEAN by @MatoTeziTanka
• Non-record: Trinity Ternary CPU v3 — Apple M1 Pro 72h, val_bpb 1.5042 #1866 — Trinity Ternary CPU v3, val_bpb 1.5042, non-record, full Track A compliance, first Apple Silicon CPU-only entry

Apologies for the noise.