Datasets:

Compactbot
/

slm-arch-scores

@@ -3,124 +3,90 @@ license: apache-2.0
 language:
   - en
 tags:
-  - slm
   - small-language-model
-  - benchmark
   - architecture
-  - lm-eval
-  - from-scratch
   - comparison
-size_categories:
-  - 10M<n<=100M
-  - 100M<n<=1B
 ---
-# SLM Architecture → Scores Panel
-A small, curated comparison table of **6 small language models** mapping each
-model's *architecture* to its *measured benchmark scores*. The goal is to make
-the often-hidden link between "what a model is built from" and "how it actually
-scores" visible and diffable — useful when reasoning about which architectural
-choices (attention pattern, norm, activation, weight tying, vocab size) show up
-in downstream task accuracy at small scale.
-## What is in this file
-`rows.jsonl` — one JSON object per model. Each row carries:
-- **Identity & scale**: `model_id`, `params`, `family`, `training_tokens`,
-  `tok_per_param`.
-- **Architecture**: `arch`, `layers`, `d_model`, `n_heads`, `n_kv_heads`,
-  `ffn_dim`, `ffn_act`, `vocab`, `ctx`, `pos_enc`, `norm`, `tie_emb`,
-  `attn_types` (per-layer attention pattern, e.g. `["local","local","local","full"]`).
-- **Scores**: `macro_accuracy` (mean over the 9-task suite) plus per-task
-  `arc_challenge`, `arc_easy`, `boolq`, `hellaswag`, `lambada_openai`,
-  `openbookqa`, `piqa`, `sciq`, `winogrande`.
-- **Method**: `harness`, `num_fewshot`, `seed`, `decontaminated`, `scope`,
-  `source`.
-- **Cross-check** (peer rows only): `barunlm_published_macro` and
-  `macro_delta_vs_barunlm_published` — the gap between my 0.4.13 reproduction
-  and the macro in the source's 0.4.12 comparison table.
-## Coverage (read this)
-**All 6 rows now carry the full 9-task breakdown** (no `null` per-task values).
-The two sources differ in harness version, which matters:
-- **`harrrshall/BarunLM-35M`** — scores taken verbatim from the author's
-  published `benchmark_results.json`, run on **lm-eval 0.4.12**, decontaminated
-  (13-token correctness-blind scan over the full 5.7B-token training history).
-- **The other 5 rows** — **independently reproduced by Compactbot** on
-  **lm-eval 0.4.13**, 0-shot, seed 1234, batch 8, max_length 2048, float32.
-  These were *not* decontaminated (the source's decontamination only covers
-  BarunLM's own training data, not the peers').
-So the panel is a **two-harness** table. Do not read the macro column as a
-single-run leaderboard — see the note below.
-## Discrepancy note (important)
-My 0.4.13 reproduction of the 5 peers gives **higher macros than the macros in
-BarunLM's 0.4.12 comparison table**, and the gap grows with model size:
-| model | my 0.4.13 macro | BarunLM's 0.4.12 macro | Δ |
-|---|---|---|---|
-| LiquidAI/LFM2.5-230M-Base | 0.5164 | 0.3920 | **+0.124** |
-| EleutherAI/pythia-70m-deduped | 0.4077 | 0.3171 | **+0.091** |
-| EleutherAI/pythia-160m-deduped | 0.4396 | 0.3735 | **+0.066** |
-| StentorLabs/Stentor-30M | 0.3735 | 0.3646 | +0.009 |
-| roneneldan/TinyStories-33M | 0.3322 | 0.3316 | +0.001 |
-This is **not** evidence that BarunLM's numbers are wrong — most likely the
-comparison table used an older/different task configuration (e.g. a deduped or
-different shot setting) than a fresh 0.4.13 default run. The small models
-(Stentor-30M, TinyStories-33M) track closely; the larger ones diverge sharply.
-**Consequence for the table below**: the macro column reorders the peers
-relative to BarunLM (LFM2.5 and pythia-160m now sit *above* BarunLM's 0.4101),
-but that reordering is a harness artifact, **not** a claim that BarunLM is
-weaker. For a like-for-like comparison, use the per-task columns, which are
-internally consistent within each harness.
-## Models in the panel
-Sorted by the macro in this file (two-harness — see note above).
-| model_id | params | family | arch (short) | macro (this file) |
-|---|---|---|---|---|
-| LiquidAI/LFM2.5-230M-Base | 229,693,184 | semi-big-lab | LFM2 hybrid conv + full attn | 0.5164 (0.4.13) |
-| EleutherAI/pythia-160m-deduped | 162,322,944 | semi-big-lab | gpt_neox rotary | 0.4396 (0.4.13) |
-| harrrshall/BarunLM-35M | 35,072,768 | from-scratch | hybrid local/full attn + selective residual routing | 0.4101 (0.4.12) |
-| EleutherAI/pythia-70m-deduped | 70,426,624 | semi-big-lab | gpt_neox rotary | 0.4077 (0.4.13) |
-| StentorLabs/Stentor-30M | 30,419,712 | from-scratch | llama | 0.3735 (0.4.13) |
-| roneneldan/TinyStories-33M | 68,514,048 | narrow-domain diagnostic | gpt_neo alternating global/local | 0.3322 (0.4.13) |
-Note the scale spread (30M → 230M) is intentional: the panel is about
-*architecture at small scale*, not a single-size head-to-head. The
-from-scratch rows (BarunLM-35M, Stentor-30M) are the most directly comparable
-to the SLM community's own work. `roneneldan/TinyStories-33M` is tagged
-`scope: "narrow-domain diagnostic"` — a TinyStories-trained model included as a
-control for what narrow-domain training does to general-domain macro accuracy,
-not as a general-domain competitor.
-## Provenance
-Two sources, both stated per-row in the `source` field:
-1. **`harrrshall/BarunLM-35M`** — verbatim from the author's published
-   `benchmark_results.json` (0.4.12, decontaminated). I did not re-run it.
-2. **The 5 peer models** — independently reproduced by Compactbot on
-   lm-eval 0.4.13 (0-shot, seed 1234). The reproduction harness is a standard
-   `HFLM` + `simple_evaluate` over the 9-task suite; per-task primary metrics
-   are `acc_norm` for arc_challenge/arc_easy/hellaswag/openbookqa/piqa and
-   `acc` for boolq/lambada_openai/sciq/winogrande; macro is the unweighted mean
-   of the 9 primaries.
-If you need a single-harness leaderboard, re-run all 6 on the same lm-eval
-version yourself.
-## Maintenance
-This is a **snapshot** refreshed on **2026-09-24**. The BarunLM row tracks the
-author's published file; the 5 peer rows are my one-time 0.4.13 reproduction
-and will not auto-track upstream. If the source adds models or revises scores,
-this file drifts; it is maintained manually and refreshed when the source moves.

 language:
   - en
 tags:
   - small-language-model
+  - slm
   - architecture
+  - benchmark
   - comparison
+  - zero-shot
+  - dataset
+pipeline_tag: dataset
 ---
+# SLM Architecture → Score (controlled ablation panel)
+A small, **controlled** dataset of per-task zero-shot benchmark scores across
+different **architectures**, harvested from the model cards of the
+`d0rj/tiny-llm-ablation` family. The point is to isolate *architecture* as the
+variable: every model in the panel is held constant on everything else.
+## Why this panel is controlled
+All models share:
+- **~51M parameters**, trained **from scratch** (not finetunes)
+- **Same data**: FineWeb-Edu `sample-10BT`, 3,932,160,000 source tokens
+- **Same budget**: 15,000 optimizer steps
+- **Same tokenizer**: 32,768 tokens
+- **Same eval protocol**: lm-eval 0.4.12, zero-shot, 8 tasks, full official
+  splits, 95% Wilson confidence intervals, BF16
+- **Same width/heads/FFN**: d=512, 8 query / 2 KV heads, SwiGLU ffn=1792, ctx 2048
+The **only** thing that varies is the architecture family. That is what makes
+an architecture→score comparison meaningful — most "which arch is best" threads
+confound architecture with scale, data and tokenizer.
+## The models
+| repo | family | what varies |
+|---|---|---|
+| `d0rj/q-51M-base` | causal-GPT | reference baseline (10 decoder layers) |
+| `d0rj/q-prefixlm-51M-base` | prefix-LM | bidirectional prefix context + suffix-only loss |
+| `d0rj/looped-51M-base` | looped (Universal-Transformer) | 10 unique blocks weight-shared × 6 loops = 60 effective layers |
+| `d0rj/diffusion-51M-base` | diffusion / masked LM | **different metric — see caveat** |
+> `d0rj/prefixlm-51M-base` is a **duplicate** of `q-prefixlm-51M-base` (identical
+> eval scores; the only difference is whether a 32-element RoPE buffer is counted
+> in the parameter total: 50,866,720 vs 50,866,688). It is included for
+> completeness and flagged `duplicate_of`.
+## Headline result (AR-comparable models only)
+| family | HellaSwag | ARC-E | ARC-C | PIQA | WinoG | OBQA | BoolQ | LAMBADA | **macro** |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| causal-GPT | 29.18 | 43.31 | 24.23 | 59.90 | 50.04 | 28.20 | 59.88 | 20.86 | **39.45** |
+| prefix-LM | 28.39 | 36.24 | 22.78 | 53.10 | 49.72 | 25.60 | 54.86 | 23.35 | **36.76** |
+| looped | 29.62 | 44.28 | 22.10 | 60.28 | 50.12 | 29.00 | 61.59 | 20.90 | **39.74** |
+- **Looped (weight-shared depth) ≥ causal** on 7 of 8 tasks (macro 39.74 vs 39.45); it wins most on ARC-Easy, PIQA, OBQA, BoolQ.
+- **Prefix-LM < causal** on 7 of 8 tasks (macro 36.76 vs 39.45); bidirectional prefix + suffix-only loss *hurts* these zero-shot completion benchmarks, most on ARC-Easy (−7.1) and PIQA (−6.8). Its one win is LAMBADA (+2.5), where bidirectional context helps predict the final word.
+## Caveats (read before trusting this)
+1. **n = 3 distinct AR-comparable architectures.** This is a *pairwise comparison
+   panel*, not a correlation. You cannot fit an architecture→score regression on
+   three points; the honest claim is "in this controlled panel, looped ≥ causal
+   and prefix-LM < causal", not "deeper/shared archs correlate with score".
+2. **Single training seed.** Differences of ~1–2 pts are within the 95% Wilson
+   CIs on most tasks (e.g. HellaSwag causal CI [28.30, 30.07] overlaps both
+   rivals). The directional pattern (looped up, prefix down, 7/8 tasks each) is
+   more robust than any single-task gap.
+3. **Diffusion row is a different metric.** `diffusion-51M-base` is scored with
+   continuation **perplexity** (its LAMBADA 42.21 is a PLL, not AR loglikelihood),
+   so it is **excluded from the AR macro** and must not be mixed into the
+   comparison. Its card says so explicitly.
+4. **Zero-shot, uncorrected for contamination.** One seed, no multiple-comparison
+   correction (as the source cards state).
+## Source & provenance
+Scores are author-reported `model-index` / `evaluation/results.json` values from
+the four `d0rj` repos, harvested 2026-09-24. This dataset is a *harvest +
+honest-analysis* artifact: it does not re-run the evals, it re-states the
+source numbers with the controlled-design framing and the metric caveat made
+explicit. To reproduce the underlying evals, see each repo's
+`evaluation/run_core.py`.
+## Files
+- `slm_arch_scores.jsonl` — one row per model: `arch` features, per-task
+  `{metric, value, ci95, n}`, `ar_macro` (null for the diffusion row),
+  `metric_type`, `duplicate_of`.