★ Current model — talk base · v3 book-pretrained 336M
The pinned status card for the owned talk model. Current champion: talk-336m — a 336M base book-pretrained on 4.92B tokens then fine-tuned on 377,700 conversation pairs, deployed and serving the live chat + the mind. Scored clean 1.00 / input-sensitive 0.96 / adjacency 0.92 on the talk_bench floors (B.20), and just got its first fluency/relevance/honesty read from the tongue gauntlet before the bridge-corpus tune. Every number here is a real, dated measurement — nothing claimed.
The pinned card. Not a writeup — a live status. How big is the model, how well did it learn, is it deployed, and can you trust it. Updated every version.
🌡️ Temperature
| Model | talk-336m (book-pretrained base — the current current pointer on SR8) |
| Version | v3 · book-pretrained champion (B.20, 2026-08-24) |
| Deployed | ✅ live on SR8 — current → talk-336m, serving the chat owned-talk phase + the mind |
| Actual size | 336M parameters (a real jump from the July 15M v2 — and the first time scale didn’t hurt) |
| Corpus | book-pretrained 4.92B tokens (14.63 tok/param, 32k BPE, EOT) → fine-tuned 16,880 steps on 377,700 conversation pairs |
| Held-out loss | 3.184 (32k-BPE chat held-out — not comparable to v2’s 2.29; different tokenizer and corpus) |
| talk_bench floors | clean 1.00 · input-sensitive 0.96 · adjacency 0.92 · stable 1.00 · ood 0.00 |
| Tongue gauntlet (before) | fluency > llama3.2:3b · relevance ~7/14 · honesty 0/6 |
| Effective size + lie-rate | earned off the sealed Gauntlet — pending |
v2 → v3: scale finally stopped hurting. For most of this project, making the owned model bigger made it worse — a reader wall that killed memorizing-in-weights. v3 is the break in that pattern: the same book-pretrain recipe scored clean 0.46 at 46M (7× fewer params, babbling), and clean 1.00 at 336M — no babble at all. It beats the prior champion on clean (1.00 vs 0.98) and input-sensitivity (0.96 vs 0.92). It loses on adjacency (0.92 vs 1.00 — it once answered a farewell with a greeting) and on out-of-distribution (0.00), which are the named next targets. It was chosen by the operator, not the floors — both checkpoints cleared every floor, so the ruler couldn’t rank them; the pick came from the qualitative gap in held-out replies (“too soon for nostalgia though, right?”) plus headroom.
Why the loss went “up” and it’s still an upgrade
The held-out loss reads 3.184, higher than v2’s 2.29 — and that is not a regression. The two numbers are measured with different tokenizers on different corpora (v2: 16k BPE on a 281k-pair set; v3: 32k BPE chat held-out after a 4.92B-token book pretrain). Cross-entropy is only comparable within one tokenizer + one eval set. The comparable numbers are the talk_bench floors above, and there v3 wins where it counts (clean, input-sensitivity) and is honestly behind where it isn’t (adjacency, ood). No cross-tokenizer loss comparison is claimed.
The tongue gauntlet — the first fluency/relevance/honesty read (before the tune)
New this version: a frozen 20-question tongue gauntlet (benchmarks/v3-tongue-gauntlet-before.md), run on the pure talk voice before the bridge-corpus tune, so there’s a dated baseline the tune has to beat. Blind-judged against llama3.2:3b:
- Fluency — the voice reads more human than the 3B: short, warm, asks follow-ups, no “as a large language model.” The floor is real and high.
- Relevance — it drifts: ~7 of 14 conversational answers are vague or off-topic. This is the axis the tune should move most, because every bridge-corpus pair is
input → a response that answers it. - Honesty — on 6 questions it cannot know (your breakfast, your car’s colour) it went 0/6: it bluffed a fake weather forecast and free-associated on the rest, against the 3B’s honest 6/6.
The read is deliberately unflattering: fluency alone was flattering the voice. A fluent non-sequitur — “pizza → blue” — is still wrong, and the gauntlet is built to catch it.
What’s next — the bridge-corpus tune (in progress)
Minting now: ~1M card-grounded conversations (the bridge — each dialogue anchored to a real card, role-played, regionally voiced). When it lands, the tongue tunes on it (free, on PC-11), and the identical 20-question gauntlet re-runs into tongue_after.json. The prediction this baseline pre-registers: relevance climbs while fluency holds. That delta — not this card — is the verdict, and it ships as a before/after when it exists.
Effective size + lie-rate are still earned, not claimed
Those two rows read pending on purpose. Effective size is earned off a reference curve — the same frozen, held-out Gauntlet run on this model and on raw models of known size (qwen 0.5B / 1.5B / 3B …); the raw size whose score equals ours is the effective size. Lie-rate — of the questions it gets wrong, what fraction it answered “I don’t know” vs. confidently lied — is the number no raw model can match. Until the sealed Gauntlet runs on v3, both read pending. No reference runs → no claim.
Versioning
Every model is a dated, versioned card. v1 was the pre-retrain babble baseline; v2 (2026-07-24) was the 15M retrained + deployed base; v3 (2026-08-24) is today’s 336M book-pretrained champion. v4 will land when the bridge-corpus tune finishes and the tongue gauntlet scores the after-picture — the first version chosen on a fluency/relevance/honesty delta, not a vibe.