Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Evaluations

Setup

Note: the numbers below were measured on the Rust engine over the full held-out validation tier (the Go and Rust engines are kept at bit-parity and decode identically).

Combined corpus of 4231 English + 4231 Chinese sentences for training; the statistics live in a unified LearnerPool (MODEL_64K_8BIT) per language.

Validation tier: 950 held-out sentences — 484 English (5.8K tokens) and 466 Chinese (5.1K tokens), never seen in training.

Substrate stats

The entire trained model — every transition statistic and open-class inventory for English and Chinese — is a 17 MB substrate on disk.

The trained substrate: 336,564 chunks, with Payload types:

  • 192,731 SPARKLE;
  • 58,129 OCTOPUS;
  • 85,704 LEARNER, that made up 2 LearnerPools;

Pool health, per 65,536-member language pool:

enzh
trained members46,33939,365
open / closed42,643 / 3,69635,036 / 4,329
total load (Σ age)401,792457,194
mean member age8.711.6
diversity margin p10 / p50 / p9051 / 128 / 25645 / 128 / 256

Held-out parse quality

Production configuration (coarse tier on), per language and combined (full-tier run on the pre-redesign substrate; the sampled tables below are current):

metricEnglishChinesemixed
full-parse99.2%99.4%99.3%
head attachment74.9%57.0%66.5%
upos95.5%87.6%91.8%
deprel77.6%64.3%71.3%
lemma92.9%92.1%92.5%
entity recall56.0%63.5%59.2%
entity precision62.5%76.2%68.1%

Entity matching is a multiset match on (type, text) between the gold entities and the surfaced ones, over parsed sentences.

Training corpus vs held-out

The same sweep per language over sampled tiers (English at 10%, Chinese at 5%), on the consolidated substrate: the training corpus bounds what the substrate retained; the held-out gap is the generalization cost.

metricen training (422)en val (65)zh training (215)zh val (24)
full-parse98.6%96.9%98.6%100%
head attachment81.9%71.1%70.1%58.5%
upos97.5%93.5%91.8%85.3%
deprel89.8%77.3%78.2%63.8%
lemma96.6%91.3%97.4%91.7%
entity recall77.3%62.9%76.6%72.7%
entity precision80.1%75.3%78.3%78.0%

The coarse tier’s contribution

Ablating the coarse-to-fine backoff on the same substrate and the same full validation tier (tier off → on), head attachment:

slicetier offtier onΔ
English (484)72.7%74.9%+2.2
Chinese (466)54.5%57.0%+2.5
mixed (950)64.1%66.5%+2.4

The gains come from trained coarse-class evidence rescuing transitions the fine-grained statistics never saw — not from structural guessing.

Decode throughput

Single decoder, sequential decode, on the same MacBook Air that trains the model in ~12 minutes (sampled tiers, consolidated substrate):

combined tokens/sval tokens/s
English21.920.8
Chinese4.14.7

The width-addressed dense posting index (no hashing on the decode-side Collect/Add) shaved ~10 ms/token off both languages on top of the pool consolidation: +23% for English; Chinese remains dominated by pool-read volume from beam churn, not per-probe cost.

Last change: , commit: ad91601