PorkiCoder Research · Tab titles

The 35M that caught Hybrid A

Scratch mid-pack GSG plus the same glue. It beats the published 35M, ties raw FLAN, and is still 0.15 behind FLAN plus centroid. Half the weights. Generate is still slow.

14 August 2026 · Holdout n=1,000 · Blind Gemini 3.5 Flash-Lite, two decimals · Sealed 1,000 unused · Follows Two recipes, one glue

Abstract

Locked Hybrid A was a 35M trained to emit Stack Overflow questions, then glued to a random-init centroid highlighter. We trained a new 35M from scratch on 4.86 million body→SO-title pairs (the mid pack), same architecture, same 2+1 glue. On one four-way Flash-Lite packet it beats locked Hybrid A 534–356 (5.65 vs 5.10) and ties plain FLAN (5.65 vs 5.64). It does not beat FLAN+centroid (5.81; 508–431). A sharper centroid (trained table, IDF, MMR) lifts FLAN to 6.11 and us to 5.96. The 35M is the lean ship: 35M parameters against FLAN-T5-small’s 77M. Latency was not closed: published Hybrid A generate is 252 ms vs FLAN’s 32 ms. We listed the levers. We did not time a faster decode.

Do not mix packets

The earlier paper’s 5.43 / 4.72 numbers are from a seven-way integer-style judge. Every number below that is two-decimal Flash-Lite with two, four, or six titles in the prompt. Absolute means move when Gemini sees more candidates. Rankings inside one packet are the science. 5.89 from a two-way is not a beat of published 5.43.

The fair packets

Same clean-dev 1,000. Same locked Hybrid A glue. Pin is scratch seed-42, step 9500 (B).

Four-way: mid GSG vs FLAN vs locked Hybrid A (old centroid)

System	Mean	Useful ≥4	Fail ≤2	Strong ≥7
FLAN + old centroid	5.81	83.1%	2.7%	36.0%
Mid GSG + old centroid (B 9500)	5.65	78.6%	3.8%	34.4%
Plain FLAN	5.64	72.1%	7.2%	40.0%
Locked Hybrid A	5.10	70.1%	5.1%	24.9%

Paired: mid vs Hybrid A 534–356. Mid vs plain FLAN 498–492. FLAN+cen vs mid 508–431. Glue cuts FLAN fail 7.2%→2.7%.

Six-way: same namers, sharper centroid

System	Mean	Useful ≥4	Fail ≤2	Strong ≥7
FLAN + v2 centroid	6.11	88.5%	0.6%	40.6%
Mid GSG + v2 centroid	5.96	85.0%	0.7%	36.8%
FLAN + old centroid	5.91	86.5%	0.5%	34.5%
Mid GSG + old centroid	5.89	85.1%	0.7%	34.5%
v2 centroid only	5.11	72.3%	1.8%	20.5%
Old centroid only	4.88	67.4%	2.1%	16.5%

v2 vs old centroid 537–438. FLAN gained +0.20 from the picker; mid GSG only +0.07. The remaining hole is 0.15. Fail is already tied.

What changed

Hybrid A trained a 35M to speak SO titles, then at decode kept at most two on-page namer words and filled to three from a random seed-42 embedding centroid. The new namer is the same 35,045,888-parameter T5 (d_model 512, 6 encoder / 3 decoder, 6,985 vocab), trained from scratch on the mid pack: 4,863,657 train pairs, one packed token cache, 12k steps, batch 256. Four arms: continue locked Hybrid A (weakest), scratch seed 42 (best peak), scratch seed 44 (same method), scratch plus a gist page-hinge (mild). Last step is not the best. Pin B step 9500. Two-way peak 5.89 vs locked Hybrid A 5.34 — that packet is only for ranking against Hybrid A.

The glue did not change. We do not call this Hybrid B. Centroid v2 swapped the random table for B 9500’s embeddings, IDF-weighted the document mean, subtracted a corpus mean, picked with MMR, and emitted in score order. Especially / Make died. Rare IDF tokens (W2k, Coplien) appeared. Cap-IDF is the next picker pass, not another 12k of question CE.

Why we are 0.15 short

Input.

Does anyone know a method to programmatically close the CD tray on Windows 2000 or higher? Open CD tray exists, but I can't seem to make it close especially under W2k.

Locked Hybrid A: Windows Tray Open
Mid GSG raw: How to close Open2m file program
Mid + glue: close Open2m Tray
FLAN: Programmatically Close CD Trace
FLAN + glue: Programmatically Close Tray

FLAN already proposes two tab words. Mid GSG still proposes a question; the glue rescues two leftovers. About 46% of glued mid-GSG words are centroid fill. More SO-title cross-entropy will not beat first-token How. The training that can close 6.11 is continue-from-B9500 on two on-page SO-title words as the target, not more questions, and not centroid words baked into the label.

Latency: listed, not measured

Published on the research Mac (PyTorch MPS, n=50): FLAN+centroid 32 ms, Hybrid A generate 252 ms, centroid-only 0.8 ms. Mid GSG is the same 35M stack. We did not rerun that clock. We did not export ONNX. We did not time a 2–4 token decode.

Levers that do not need a new quality story:

Greedy, 2–4 new tokens, stop at two content words. Glue still fills. This is also the quality experiment (constrained decode).
No beam, no best-of-k at ship. Extra samples buy judge points and blow the 32 ms budget.
Gist / shuffle / generic hinges are train-only. Infer is one forward generate.
Centroid is 0.8 ms on CPU. It is not the hitch.
Do not drop this 3-decoder model into the live 12.7 worker. That worker hardcodes two decoder layers.

Size is already the win versus FLAN: 35M vs 77M. Speed is not. Quality first, then cut decode length, then ONNX INT8 on a new worker.

What failed, again

Continuing the published Hybrid A namer at 1e-4. Weakest of the four trainers.
Extractive 48k with document-order fill. Spoke stickers, memorized Asp Net.
Gemini-title CE from scratch, span-pretrain, 70M as a ship path. Closed.
Three boxes tokenizing the same jsonl. One packer, then trainers from the pack only.
Exact match, retrieval, and all_onpage as promotion metrics.

What to ship, and what not to claim

Phone-class slot: mid GSG 35M + locked glue, checkpoint B 9500. 16 GB Mac slot: still FLAN + centroid until a same-packet mid GSG row beats 6.11 (or 5.81 on the old centroid). Do not market this as “our 35M beat Google.” It beat our last 35M and tied raw FLAN.

Weights live in the research checkout under output/gsg_scaleup_20260814/keepers/B/checkpoints/step_9500/. How to run them: SHIP_35M_MID_GSG.md in this repo. The four Ada droplets used for training were deleted after the keepers landed. brisbane-copilot and pigwise were not.