udgamlabs/small-model-board-scores

UDGAM LABS · Dataset · Per-item scores for 27 language models up to 400M parameters

Small Model Board scores

Whether each of 27 small language models got each benchmark item right, for every item of eight standard benchmarks: 1,152,450 per-item scores, plus the aggregate scores of all 27 models. These are the numbers behind the Small Model Board. Every model ran at a pinned revision on one harness, with the same settings, on 19 October 2026. The dataset holds no question or answer text: items are identified by their index, their id and lm-eval's SHA-256 item hash, so you can join them back to the original benchmarks yourself.

Use it to check a number on the board, re-score with another rule (for example length-normalised cloze), compute paired statistics between two models on exactly the same items, or find the items every small model gets wrong.

Quick start

from datasets import load_dataset

scores = load_dataset("udgamlabs/small-model-board-scores", "scores", split="train").to_pandas()
models = load_dataset("udgamlabs/small-model-board-scores", "models", split="train").to_pandas()

# MMLU cloze minus letter for one model, paired by question
m = scores[scores.model == "HuggingFaceTB/SmolLM2-135M"]
mmlu = m[m.benchmark.isin(["mmlu_letter", "mmlu_cloze"])].pivot_table(
    index=["subtask", "item_index"], columns="benchmark", values="correct")
print(100 * (mmlu.mmlu_cloze - mmlu.mmlu_letter).mean())   # +5.40

Results

The aggregate table is the models config and the board. Averaging correct over a model's items in scores gives the board's number for that benchmark exactly. Before release we recomputed every one of the 1,244,646 item results from the raw log-likelihoods in our lm-eval logs: all equal the per-item values lm-eval logged, and all 216 benchmark averages equal lm-eval's own results files.

Where it falls short

• These are our own runs, not a certified leaderboard: one harness version (lm-eval 0.4.13), one GPU type, bf16 at batch size 64. Other harnesses, prompts or hardware give different numbers, and bf16 scoring can flip a few close items from run to run.
• Instruct and chat models were scored without their chat templates, as plain text completers; that is not how they are meant to be used.
• Q-164M has a built-in calculator, so its GSM8K score is tool-assisted.
• MobileLLM-R1-140M-base and nltower-100m-base have summary scores only, with no per-question rows. The per-question chosen and pred values are model outputs. MobileLLM-R1 is under the FAIR Noncommercial Research License, and nltower-100m-base's repo carries no licence text, so we do not redistribute their outputs. Their benchmark scores are in models and on the board (see "Per-model notices").
• Dated 19 October 2026: newer models, for example Boris-2-Preview-2 (21 October), are not in it.
• Letter-format MMLU sits at or below the 26.89% of always answering "D" for 23 of the 27 models, so at this size it mostly measures whether a model can answer in letter format, not what it knows.
• We did not check any model for benchmark contamination.

What is inside

scores: one row per model × benchmark × item (25 models × 46,098 items; the two summary-only models above are left out for licence reasons).

Column	Meaning
model	Hugging Face repo id
revision	the pinned commit that was evaluated
benchmark	arc_easy, arc_challenge, hellaswag, piqa, winogrande, mmlu_letter, mmlu_cloze, gsm8k
subtask	MMLU subject (null otherwise)
item_index	0-based row of the item in the split lm-eval reads (lm-eval's doc_id)
item_id	the benchmark's own id for ARC (for example Mercury_417466); subject/index for MMLU; the index otherwise
metric	the metric the board uses for this benchmark: acc, acc_norm or exact_match
correct	1 if right under metric, else 0
chosen	the option the model picked under metric (0-based; null for GSM8K)
acc, pred	right/wrong and picked option by highest log-likelihood
acc_norm, pred_norm	the same after dividing by the option's length in characters (null where lm-eval does not compute it: WinoGrande and MMLU letter)

items: one row per benchmark item (46,098): benchmark, subtask, item_index, item_id, split, doc_hash (lm-eval's SHA-256 of the item, for checking a join) and n_choices.

models: one row per model (27): model, revision, evaluated_as, group (≤ 200M or 200M–400M parameters as loaded), type (what the model's own card says: base, instruct or not stated), params, by_udgam_labs, add_bos_token, remote_code, license_tag, per_item_rows, and the score of every benchmark in percent, the 5-task average (avg5) and cloze_minus_letter.

Udgam Labs' own five models are in the table like any other: udgamlabs/Laghu-Base, udgamlabs/Gargi-Base, udgamlabs/Gargi, udgamlabs/Udgam-384M-Chat and udgamlabs/Udgam-384M-Base-4K. Laghu-Base was evaluated from our training repository; its weights, config and tokenizer files are byte-identical (SHA-256) to udgamlabs/Laghu-Base @ 479e6926, which is the revision listed here.

How it was made

• lm-eval 0.4.13 with --log_samples, transformers 5.19.0, torch 2.14.1, one NVIDIA RTX PRO 6000 Blackwell, bf16, batch size 64, seed 1234, no chat template, 18–19 October 2026 (UTC).
• Zero-shot ARC-Easy (acc), ARC-Challenge (acc_norm), HellaSwag (acc_norm), PIQA (acc_norm), WinoGrande (acc), MMLU in letter format (mmlu, acc) and cloze format (mmlu_continuation, acc); GSM8K 5-shot, exact match with lm-eval's flexible-extract filter. lm-eval's default task configurations.
• Udgam Labs models were run with add_bos_token=True, as they were trained with a <bos> token; every other model with its own tokenizer defaults.
• Ten models need their own modeling code. We read that code without running it, checked it for network, shell, exec/eval, pickle, download and file-write calls, and pinned the reviewed commit before running it on a machine that held no credentials.
• Not runnable in the harness, so not included: Nikity/lille-130m-base (no modeling code in the repo), Maggio33/GoLLeM-v5-128M (no model_type in its config) and Cactus-Compute/needle3 (a tool-calling model with no LM head).

Benchmark items

Benchmark	Source and revision	Split, items	Licence as stated by the source
ARC-Easy, ARC-Challenge	allenai/ai2_arc @ 210d026f	test, 2,376 / 1,172	CC-BY-SA-4.0
HellaSwag	Rowan/hellaswag @ 218ec52e	validation, 10,042	MIT
PIQA	baber/piqa @ 142f6d73 (lm-eval's copy of ybisk/piqa)	validation, 1,838	Academic Free License v3.0 on the authors' page; "unknown" on the ybisk/piqa card
WinoGrande (winogrande_xl)	allenai/winogrande @ 01e74176	validation, 1,267	data CC-BY, code Apache-2.0 (the authors' README)
MMLU	cais/mmlu @ c30699e8	test, 14,042 (in both formats)	MIT
GSM8K	openai/gsm8k @ 740312ad	test, 1,319	MIT

Each revision is the one the source served during the run: by the Hub's last-modified dates, the newest change to any of them was in March 2026. Why no text: ARC is share-alike, PIQA's licence is stated differently in different places, and copying test questions to one more place makes it easier for them to leak into training data. Ids, indices and hashes are enough to join the scores back to the sources.

Per-model notices

Per-item chosen and pred values are model outputs.

• MobileLLM-R1-140M-base (Meta): reported for noncommercial research under the FAIR Noncommercial Research License; we acknowledge Meta's materials. Aggregate scores only; no per-item rows.
• nltower-100m-base: the model repo carries no licence text (its card's licence field names "climbmix-dolmino-pretraining-terms" without a link). Aggregate scores only; no per-item rows.
• Falcon-H1-Tiny-90M-Base and -Instruct: made by the Technology Innovation Institute (TII), licensed under the TII Falcon License and the Falcon Acceptable Use Policy, https://falconllm.tii.ae/falcon-terms-and-conditions.html.
• gemma-3-270m: Gemma is provided under and subject to the Gemma Terms of Use, https://ai.google.dev/gemma/terms.
• LFM2.5-230M (Liquid AI): licensed under the LFM Open License v1.0.
• BananaMind-2-Pro: licensed under the BananaMind Community License 1.0.
• All other models carry an Apache-2.0 licence tag; license_tag in models gives each one.

Intended use and limitations

For research on small language models: checking the board, re-scoring, paired comparisons and error analysis. Not a certified ranking, and not a substitute for evaluating a model on your own task. Do not train on the benchmark items these scores point to.

Licence

• Our scores, tables and annotations (every per-item score, the aggregates and this card): CC-BY-4.0, © 2026 Udgam Labs LLP.
• Other models' outputs (chosen, pred, pred_norm) are released under CC-BY-4.0 as far as we hold rights in them, together with the notices above.
• Benchmark items are not included; they keep their upstream licences (table above).
• MobileLLM aggregates are reported for noncommercial research under the FAIR Noncommercial Research License.

Citation

@misc{udgamlabs2026smallmodelboard,
  title        = {Small Model Board scores},
  author       = {{Udgam Labs}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/udgamlabs/small-model-board-scores}}
}

Small Model Board by Udgam Labs · AI research from first principles

---

About Udgam Labs. Udgam Labs is an independent AI research lab. We work from first principles: before we build, we understand why things work, whether we are training a model, fine-tuning one or applying AI to a new problem. Udgam (उद्गम) is Sanskrit for origin. udgamlabs.com