PorkiCoder Research · Tab titles

Two recipes, one glue

16 GB Macs get FLAN plus our highlighter. Phone-class chips get the 35M we actually trained. The same A18 Pro sits in a $599 Neo and in an iPhone; that is the split, not Air versus Pro.

14 August 2026 · Holdout n=1,000 · Blind Gemini 3.5 Flash-Lite · Sealed 1,000 unused

Where each recipe lives

A18 Pro is the iPhone 16 Pro SoC. Neo is that phone in a clamshell. An M4/M5 Air is the same kind of silicon as a Pro, with a 16 GB base, not a different planet. Do not put Hybrid A on the Air and FLAN on the Pro. That split is 8 GB phone-class vs 16 GB Mac. The packaged bundle sizes below decide the final cut; the Hybrid A slim figure is an estimate until the export exists.

The holdout table

Same blinded packet. Latency is single-title on the research Mac (PyTorch MPS generate, CPU centroid, n=50). Combiners run sequentially.

Useful = Gemini ≥4. Fail = Gemini ≤2. FLAN+centroid vs raw FLAN, paired wins: 427–374. vs Hybrid A: 449–246. Sealed final 1,000 was not scored. *Slim = GSG weights plus embed-only centroid; measured research checkpoint is 272 MiB, the slim export has not been built yet.

Size, speed, useful, fail

X is milliseconds per title. Y is the Gemini rate. Green fill is useful. Red ring is fail. Bubble area is disk.

FLAN sits left (fast) and high. Our 35M generate sits near 250 ms. The triple is the slowest for a tie with Hybrid A. Centroid is a green speck on the y-axis.

What we thought we were building

PorkiCoder needs a local tab title: a few words so you can find the session later. The 12.7M student that ships in the app today collapsed to brand priors (Java EE, ASP NET MVC). That model has a history: it is the distilled student whose training quirks we dissected in our first paper, The Sniff Test, where one lucky embedding transplant looked like a mechanism until 31 matched runs said otherwise. This campaign started as its replacement.

We spent a GPU day on a 35M T5: technical span-corruption pretraining, then PEGASUS-style GSG (task body → original Stack Overflow title). That model learned a title dialect (How to a varchar field in Active), not a gist. Exact match to the SO title was never the product metric.

The thing that actually wrote stickers was dumb: take the three body words whose embeddings sit nearest the body’s centroid, keep document order. Sixteen mebibytes if you throw away the unused transformer.

How the pieces were formed

Data

Every system reads the same serialization the app uses: title: agent=… task=… with the task body from a public Stack Overflow dump. Technical-language pretraining drew 250M tokens from shards 18–57 of that dump. Evaluation lives in so_clean_holdout_v1: 2,000 newly pulled tasks from shards 0–17 only, with all 49,336 previously used title-training and validation IDs excluded, split by hash sort before any title training. Half is the reusable clean dev set of 1,000 that every number on this page comes from; the other 1,000 are sealed behind a recorded SHA-256 and have never been generated on, labeled, judged, or read.

The models

The 35M is a 35,045,888-parameter T5: d_model 512, d_ff 2048, 6 encoder and 3 decoder layers, a 6,985-token compact vocabulary, tied embeddings. GSG continued from the span-corruption pretrain to step 3,000 on body-to-title pairs. The FLAN row is not untouched stock: it is FLAN-T5-small (77M, Apache-2.0) after the title fine-tune from run 5 of our earlier distillation campaign, where it served as the capacity reference. Stock google/flan-t5-small is the same family; the shipping bundle pins one exact checkpoint by SHA-256.

The judge

Gemini 3.5 Flash-Lite scores blind: it sees the task and anonymous candidate titles in randomized order, reference labels hidden, and rates 1–10 for whether the title would help you find that tab again. Useful is 4 or above, fail is 2 or below. An earlier hand-scored 90-task packet ranked systems the same way but with harsher absolute numbers (Hybrid A 56% useful, 12.7M 13%), so trust the ranking and the gaps, not any single percentage. Latency is the median of 50 single-title runs, generate on Apple MPS, centroid on CPU.

What the centroid actually is

Not magic, and honestly, not even trained. The highlighter owns a 512-dimensional embedding table for the 6,985-token vocabulary. Each content word of the task body (stop words and short tokens dropped) becomes the mean of its subword rows, normalized. The centroid is the normalized mean of those word vectors. The title is the two or three words with the highest cosine to the centroid, printed in document order, capitalized.

The table is the seed-42 random initialization of the 35M, frozen before any gradient ever touched it. We tried training embeddings with a cosine objective and they collapsed; the untrained table as a filter is what survived. So the mechanism is closer to a random projection voting on which words share the page’s token statistics than to learned semantics. It costs 13.6 MiB of matrix plus 2.3 MiB of tokenizer, runs in 0.8 ms, and beats the 12.7M neural incumbent by 1.2 judge points. That result is the quiet embarrassment of this whole campaign, and we are shipping it.

The glue

Hybrid A takes up to two content words from the 35M GSG title that actually appear in the task body (substring matches count for words longer than three characters), deduplicates, then fills to three with centroid nouns. If the namer contributes nothing that survives the in-body check, the centroid words stand alone. The same frozen rule with FLAN as the first source is the 16 GB Mac recipe. The stop list, token regex, and 2+1 fill rule are locked; changing any of them requires a new scored packet.

Dead ends, so nobody repeats them

• Title cross-entropy from scratch on 50k FLAN-labeled tasks collapsed to one title (Fix XML Files), twice, including once with a Kimi-reviewed recipe. Axis closed.
• The span-pretrained 35M learned language, not titles: a blank source beat the correct source on NLL.
• Cosine as a loss collapsed embeddings. Cosine as a filter is fine.
• Instruction-continuing GSG on extractive FLAN titles gained nothing over locked Hybrid A.
• Adding the 12.7M into the glue tied Hybrid A at double the latency and sometimes pasted EE into titles.

Labels, with the input

can anybody recommend some really good resources for how to get Apache authenticating users with Kerberos. Background reading on Kerberos would also be useful Thanks Peter

I need a C# profiler. Although I'm not averse to paying for one, something which is free or at least with a trial version would be ideal since it takes time to raise a purchase order. Any recommendations?

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.

I'm familiar with SQL Server Indexed Views (or Oracle Materialized Views), we use them in our OLAP applications. They have the really cool feature of being able to usurp an execution plan…

FLAN names the job. Centroid highlights nouns that are already on the page. 12.7 invents a stack. Hybrid A is a highlighter with an occasional stack word from GSG. FLAN+centroid is a namer whose off-page junk gets replaced by nouns.

Appendix A: the whole packet, in numbers

Summary statistics over all 1,000 clean-dev tasks, one blinded seven-way Gemini call per task. Gold is a score of 8 or higher. Unique titles counts distinct strings across the 1,000. Top-title mass is the share of the single most repeated title, the collapse detector that killed earlier recipes. No system emitted an empty title. W–L is paired wins and losses against FLAN+centroid on the same task; the remainder is ties.

FLAN keeps the gold medals (25.2% of its titles score 8 or higher, against 18.4% after the glue) yet still loses the pairing 374–427. The glue trades peaks for consistency: median 6, fail under 10%. The 12.7M’s median score is 2.

Appendix B: thirty tasks, six labels each

The first 30 rows of the scored packet in file order. No selection, no reordering. The number in parentheses is that title’s blind Gemini score on the task. Triple is the 12.7 + GSG + centroid stack.

Cost, honestly

FLAN-T5-small is T5-small trained on C4 at Google, then instruction-tuned on more than 1,000 FLAN tasks on TPU v3/v4 pods. The 296 MiB file is that entire stack, Apache-2.0. It is not a weekend 77M. Our 35M pretrain, GSG, and all judging ran on one rented RTX 6000 Ada at $1.57/hr plus a few dollars of Gemini calls: tens of dollars total. Hybrid A is close for a pretrain. It is not a peer of FLAN as a recipe. It is the model that fits next to an A18 without pretending we have a Pro’s memory. The 16 GB Mac borrows the summarizer and keeps the highlighter. Phone-class silicon keeps the highlighter and the 35M we trained.

One rule changed to say this out loud: our evaluation preregistration stated that FLAN was a benchmark only and this project would ship only its own model. That rule is overridden by an explicit owner decision, recorded here. The shipped contribution is the centroid table plus the in-body glue, which cut FLAN’s fail rate from 18% to 10% and lifted useful from 68% to 77%. We are not marketing a from-scratch 35M win, because there is not one.

What we are not claiming

• We did not beat FLAN by training a better 35M.
• The centroid table is untrained. No claim of learned semantics, only a filter that works.
• The FLAN numbers come from our run-5 title-tuned descendant of FLAN-T5-small, not the untouched Hugging Face file. The family is the same; the exact shipped checkpoint gets pinned by hash.
• We did not score the sealed 1,000. Clean-dev 1,000 is continuity, not the final gate.
• Gemini is more generous than our hand rubric. Rankings held across packets; absolute percents move when FLAN sits in the same row.
• The Hybrid A slim bundle size and the phone-class memory behavior are projections until the export is built and measured on device.

Ship path

Two bundles, one combiner, pick the namer by RAM. Not a third model.

• resources/tab-namer/flan-centroid-v1/: default on 16 GB+ Mac.
• resources/tab-namer/hybrid-a-v1/: slim 35M GSG + centroid; default on A18 / 8 GB; fallback under memory pressure.
• Centroid-only if the OS is jetsamming.
• Directory name if nothing validates.

New workers. Do not reuse the 12.7 ONNX graph (2 decoder layers, 6,985-id vocab) for FLAN, which needs its full 32k vocabulary and 8+8 layers. Sanitize into the 1–3 word [A-Za-z0-9 ] contract before rejecting, and fall back to centroid if sanitizing empties the title. First release runs in shadow mode: compute, log, keep showing the current title. Then switch on. Details: FLAN_CENTROID_HANDOFF.md in the PorkiCoder checkout.