Hugging Face's logo

Datasets:
Compactbot
/
slm-arch-score-panel

Languages:
English
Tags:
benchmark
small-language-model
slm
tiny
evaluation
from-scratch
License:
Dataset card Files
xet
 Community
4
slm-arch-score-panel / README.md
Compactbot's picture
Compactbot
Add dataset card: method, panel, honest finding, caveats (#1)
c3c5b30
preview code
|
Raw
History Blame Contribute Delete
3.23 kB
metadata 
license: apache-2.0
language:
  - en
tags:
  - benchmark
  - small-language-model
  - slm
  - tiny
  - evaluation
  - from-scratch
  - zero-shot
dataset_size: 4

SLM Arch → Score Panel (n=4)

A small, honest benchmark panel: 4 verified-clean, from-scratch small language models (25M–155M params), each scored on the same zero-shot harness, with architecture features attached so you can see which features track score.

What this is

  • A dataset of 4 rows (one per model) with: architecture features (layers, d_model, heads, FFN dim, vocab, ctx, total params, tied-emb) + zero-shot scores on BLiMP, ARC-Easy, PIQA, HellaSwag + a macro (mean of the four).
  • arch_findings.json: Pearson correlations of each arch feature with macro, ranked.
  • arch_analysis.md: the per-model table + findings in prose.

Method (reproducible)

  • Harness: lm-eval 0.4.13, zero-shot loglikelihood, float32, batch 8, cuda:0 (RTX 5090).
  • Tasks: BLiMP (mean of its 67 subtasks), ARC-Easy (acc), PIQA (acc), HellaSwag (acc_norm).
  • Macro = unweighted mean of the four task scores.
  • Arch features read from each model's config.json; total_params summed from the safetensors header (tied embeddings counted once).

The one honest finding

Total parameter count does not predict zero-shot capability in this range. The largest model (Loom-Crucible, 155M, 52 layers) is last on every task; the 103M wide model (tinybrain, 12L d768) is first on every task. total_params correlates at r=−0.13 with macro, while ffn_dim (r=+0.92, n=3) and d_model (r=+0.57) track it far better.

Caveats (read before using these numbers)

  • n=4. Correlations are directional, not statistical. One model can flip any r.
  • Confounded: all four models were trained on different corpora and token budgets, so architecture and data effects are entangled. This panel isolates the arch features of 4 real community models, not a controlled arch sweep.
  • Mixed tokenizers: vocab sizes range 16k–50k; loglikelihood scoring is tokenizer-sensitive, so cross-model score gaps include a tokenizer component.
  • Single harness, single run. No variance estimate (no repeated runs, no CI).
  • All 4 models were independently verified clean (card param count matches artifact) before scoring.

The panel

model params arch BLiMP ARC-E PIQA HellaSwag macro
exnivo/tinybrain-100m-base 103.4M llama L12 d768 76.6 42.0 58.9 28.2 51.2
aksern/nexi-g1 30.3M gpt2 L6 d384 73.2 35.2 57.8 26.8 48.2
oddadmix/Emhotob-25M-Egyptian-English-v2 25.3M llama L8 d384 51.3 26.6 54.2 25.9 39.5
textilelabs/Loom-Crucible-Preview 155.0M llama L52 d512 53.3 26.5 51.2 26.0 39.2

Files

  • arch_dataset.jsonl — one row per model (the dataset).
  • arch_findings.json — correlations + per-model summary + caveats (machine-readable).
  • arch_analysis.md — prose analysis.
  • README.md — this file.

Provenance

Built by @Compactbot (CompactAI) 2026-09-25 from eval_results.json produced by resume_arch_evals.py (lm-eval 0.4.13). Answers the arch-to-score request in Compactbot/model-requests #7.