PorkiCoder Research · Tab titles
176k Flash-Lite CE
A quantity test after 6t. Four LRs, one pack, one Terminal packet. Ship stays 6t_argmax.
Still ship this
6t_argmax, sha 2ea02c13…. Same weights as production porkr1/resources/tab-namer/6t-argmax-centroid-v1. Beam-4 two-word + Hybrid A + centroid v2. This 176k run does not replace it.
Two boards
This page is Terminal-board only (300 terminal-dev). SO-board is not mixed in. Holdout 200 and sealed SO 1000 stay sealed. Product metric is blinded Gemini 3.5 Flash-Lite on Hybrid-A-glued titles, two decimals. CE has no ship weight.
Abstract
We stopped a 1M Flash-Lite labeling job at 219,800 raw rows (~$16) and trained on 175,642 vocab-legal titles. Four 35M continues (A 2e-5, B 5e-5, C pin+2e-5, D 1e-5) ran 8,000 steps on four Ada 6000s. CE fell 6.2 → ~1.5 (B to 0.91). Unique stayed ~290/300. On the first glue-only packet, A was 6.19 vs 6t 6.59 vs FLAN 7.05. After adding 6t’s beam-4 two-word decode, A is 6.05 vs 6t 6.16 vs FLAN 6.62 (Δ A−6t −0.112). Hottest CE (B) is still last. Quantity of teacher-CE did not beat the small judged set or FLAN.
What we trained on
GH Archive agent/PR pages, labeled by gemini-3.5-flash-lite (2–3 words, on-page overlap ≥ 0.5, 7k vocab-legal, ≤16 target tokens). 6t was judged-SFT on 2,040 reward≥6 titles. This pack is ~86× more rows and not pickier.
IDF 2-word stickers were tried first on the same 1M pages and produced gibberish. Flash-Lite labels replaced them. Next paid regen must say max 3 short common words — see required-specs.md.
Learning rates
Same pack, same 35M, 8k steps, batch 256. Notes written at 18:20 PDT before the 8k snap, confirmed by the packets below.
| Arm | Init | LR | Final CE | What we learned |
|---|---|---|---|---|
| A | 6t | 2e-5 | 1.55 | Workhorse. On-page held. Closest to 6t after beam-4. |
| B | 6t | 5e-5 | 0.91 | Prettiest curve. Raw on-page fell 0.32→0.18. Worst judge. |
| C | pin B-9500 | 2e-5 | 1.54 | A’s twin. LR dominates init on this teacher. |
| D | 6t | 1e-5 | 1.92 | Slow spare. Spikes were batch noise, not a reject. |
Packet 1 — glue only (unfair to 6t)
Trainer beam + Hybrid A. 6t already had production beam-4. n=300, same judge, same centroid.
| Arm | Mean | Useful ≥4 | Fail ≤2 | vs 6t |
|---|---|---|---|---|
| FLAN + centroid | 7.05 | 93.3% | 1.7% | — |
| 6t_argmax (ship) | 6.59 | 92.3% | 2.3% | — |
| A 2e-5 | 6.19 | 85.3% | 2.7% | −0.399 (101–80–119) |
| D 1e-5 | 6.13 | 86.0% | 3.0% | −0.461 |
| C pin 2e-5 | 6.12 | 86.7% | 3.0% | −0.467 |
| B 5e-5 | 6.06 | 84.7% | 3.3% | −0.534 |
Do not compare these absolute means to packet 2. New judge draw. Paired Δ inside a packet is the fair number.
Packet 2 — full 6t stack (apples to apples)
A–D re-decoded with eval_snapshot_router.py: beam-4 two-word, Hybrid A bounded 32, centroid v2. Same 300, same 6t routed file, same FLAN Hybrid A, one blinded Flash-Lite packet.
| Arm | Mean | Useful ≥4 | Fail ≤2 | vs 6t |
|---|---|---|---|---|
| FLAN + centroid | 6.62 | 89.0% | 3.0% | — |
| 6t_argmax (ship) | 6.16 | 87.3% | 3.7% | — |
| A 2e-5 | 6.05 | 85.0% | 3.0% | −0.112 (73–149–78) |
| B 5e-5 | 6.02 | 86.7% | 3.0% | −0.135 |
| C pin 2e-5 | 6.00 | 85.3% | 3.0% | −0.162 |
| D 1e-5 | 5.99 | 86.0% | 3.0% | −0.164 |
Beam-4 closed most of the fake gap (−0.40 → −0.11). They still lose. B’s CE win is still a judge loss.
Learnings
- Copying Flash-Lite tokens is not inheriting Flash-Lite taste. Transitivity fails at 7k vocab + beam-4.
- Hottest LR wins the chart and loses the page.
- IDF stickers are not a teacher. Flash-Lite 2–3 word labels are a real teacher and still not enough CE.
- mkdir remote
scripts/before rsync or env-ready boxes sit idle (required-specs.md). - Do not scale this CE recipe to 1M as a 6t replacement. Change the teacher filter or the objective first.
Do not put back on the plate
- Promote B because CE is 0.91.
- Mix SO-board and Terminal-board in one table.
- Regenerate the paid 219,800 labels unless the owner asks.
- Delete brisbane-copilot or pigwise with campaign Adas.
Artifacts
LR notes: 2026-08-16_1820PDT_g300k_learning_rate_findings.md.
Packets: output/g300k_eval/packet.summary.json, output/g300k_eval/beam4/packet.summary.json.
Labels: data/derived/terminal/agent_1m_gemini/shards/ (Gitea).
Weights: local output/g300k_train/*/final (not in git).
Production 6t: porkr1/resources/tab-namer/6t-argmax-centroid-v1.