slm-arch-score-panel / arch_analysis.md
Compactbot's picture
Prose analysis: per-model table + arch-feature correlations (#3)
b8e04d1
|
Raw
History Blame Contribute Delete
1.64 kB

Arch -> Score: what the numbers say

Panel: 4 from-scratch SLMs, all verified-clean, all in the small-model community.

Every model scored on the SAME harness (zero-shot loglikelihood, lm-eval 0.4.13, float32, batch 8, cuda:0): BLiMP (mean of subtasks), ARC-Easy, PIQA, HellaSwag (acc_norm). Macro = mean of the four.

Per-model

model params arch BLiMP ARC-E PIQA HellaSwag macro
exnivo/tinybrain-100m-base 103,385,856 llama L12 d768 76.6 42.0 58.9 28.2 51.2
aksern/nexi-g1 30,339,456 gpt2 L6 d384 73.2 35.2 57.8 26.8 48.2
oddadmix/Emhotob-25M-Egyptian-English-v2 25,271,424 llama L8 d384 51.3 26.6 54.2 25.9 39.5
textilelabs/Loom-Crucible-Preview 154,980,864 llama L52 d512 53.3 26.5 51.2 26.0 39.2

Arch feature -> macro correlations (Pearson r)

feature r n
ffn_dim +0.921 3
d_model +0.567 4
n_heads +0.567 4
n_layers -0.538 4
vocab_size +0.365 4
max_ctx +0.154 4
total_params -0.132 4

Best-arch read

Highest macro: exnivo/tinybrain-100m-base (51.2). Its arch: {"total_params": 103385856, "n_layers": 12, "d_model": 768, "n_heads": 12, "ffn_dim": 2048, "vocab_size": 24000, "max_ctx": 2048}

Honest caveat: with 4 models these correlations are directional, not statistical. The strongest signal is that parameter count and depth track macro, but every model here was trained on a different corpus and token budget, so architecture and data are confounded. This panel is a first consistent measurement, not a verdict.