udgamlabs/laghu-evals

UDGAM LABS · Dataset · Evaluation outputs for Laghu (152M) and 37 other models

Laghu evaluation outputs

Every evaluation behind Laghu's model card and technical report, with the output and score of every single item, so anyone can check a number, re-score it with a different rule, or compare a new model on exactly the same items. Laghu is Udgam Labs' 152M-parameter chat model. The dataset holds four parts:

• Chat suite (pre-registered, measured 19 October 2026): Laghu and 14 other chat models answering GSM8K (1,319), MATH-500 levels 1-3 (238), IFEval (541), 214 two-turn follow-ups and 324 real-user chat prompts, plus the scorer-check runs (two more models). 58,125 generations and 53,367 per-item scores for Laghu and 13 of the rivals, and 1,944 pairwise chat-quality verdicts from an LLM judge.
• Base suite (lm-eval 0.4.13, 19 October 2026): Laghu-Base and 26 other small base and chat models on ARC-Easy, ARC-Challenge, HellaSwag, PIQA, WinoGrande, MMLU (letter and cloze formats) and GSM8K 5-shot. 25 models have per-item log-likelihoods and generations (about 47,400 rows each).
• Held-out bits per byte: per-document results for Laghu-Base and five other base models on 1,695 documents written in September and October 2026, plus two held-out sets from our own data lineage.
• Long context: the checks of Laghu's 32K stretch (needle recall, copy accuracy, bits per byte by position) and the base suite before and after it.

Model: udgamlabs/Laghu · Code and scorers: github.com/udgamlabs/laghu · udgamlabs.com

Quick start

from datasets import load_dataset
prompts = load_dataset("udgamlabs/laghu-evals", "chat_prompts", split="test")
verdicts = load_dataset("udgamlabs/laghu-evals", "judge_verdicts", split="smollm2_135m_instruct")

Generations, scores and lm-eval rows are gzip-compressed JSON Lines, one file per model and task:

import gzip, json
from huggingface_hub import hf_hub_download
p = hf_hub_download("udgamlabs/laghu-evals", "chat_suite/generations/laghu/vllm/gsm8k__on__greedy.jsonl.gz", repo_type="dataset")
rows = [json.loads(l) for l in gzip.open(p, "rt")]
print(rows[0]["id"], rows[0]["raw"][:300])

Some benchmarks ship as ids and scores only (see "Benchmark items"). join_items.py puts their text back from pinned public revisions and checks every row against lm-eval's own hash:

huggingface-cli download udgamlabs/laghu-evals --repo-type dataset --local-dir laghu-evals
python laghu-evals/join_items.py laghu-evals --out laghu-evals-joined            # needs huggingface_hub, pyarrow

Results

Every number below is [MEASURED] by us, recomputed from the per-item files in this dataset (the build re-scores them and compares with the published tables; see "How it was made").

Chat suite, pre-registered comparison (vLLM, bf16, each model's own chat template, the same items and scorer for every model; Laghu with thinking off unless stated). Source: tables/CLAIMS_2B.md, tables/TABLES_full.md.

Category (metric)	Laghu	SmolLM2-135M-Instruct	Falcon-H1-Tiny-90M-Instruct	MobileLLM-R1.5-140M
Maths with thinking (GSM8K test, 1,319, strict: a reply cut off by the cap counts as wrong)	7.9%	1.8%	23.3%	0.0%
Multi-turn (follow-up right, 144 checkable items)	41.7%	37.5%	54.9%	3.5%
Chat quality: Laghu's length-controlled win rate vs the rival (324 prompts × 2 orders; judge DeepSeek-V4.1-Flash)	—	51.0% [46.5, 55.5]	28.7% [24.0, 33.6]	57.3% [27.0, 76.2] (raw 91.0%, length-confounded)

MobileLLM results are reported for noncommercial research under the FAIR Noncommercial Research License; we acknowledge Meta's materials.

After Holm correction over the nine comparisons, Laghu wins no category against all three rivals and trails Falcon-H1-Tiny-90M-Instruct on all three. Against our own 384M Gargi, Laghu shows no significant difference on chat quality (45.9% win rate, p = 0.06) or multi-turn (41.7 vs 49.3, p = 0.12), and trails it on maths (7.9 vs 35.6).

Base suite (lm-eval 0.4.13, bf16, zero-shot except GSM8K 5-shot). Source: tables/BASE_RESULTS_CUDA.md. Laghu-Base averages 49.9 on ARC-Easy, ARC-Challenge, HellaSwag, PIQA and WinoGrande: 5th of the 19 models of 200M parameters or fewer that we measured. The best of them, GPT-X3-150M, averages 53.0; SmolLM2-135M averages 51.8.

New text (bits per byte, lower is better, 1,695 documents written 1 September to 15 October 2026). Source: bpb/summary.json. Laghu-Base 0.919; SmolLM2-135M 0.941; Falcon-H1-Tiny-90M-Base 0.979; Gargi-Base (384M) 0.914.

Long context. After the 32K stretch, Laghu recalls a 5-digit code in 100 / 100 / 99 / 96% of 100 trials at 4K / 8K / 12K / 16K tokens, and in none at 24K or 32K. Source: long_context/RESULTS_32K.md.

Where it falls short

About Laghu, as these files show it:

• Maths: 7.9% on GSM8K with thinking (strict), against 23.3% for Falcon-H1-Tiny-90M-Instruct and 35.6% for Gargi.
• Instruction following: IFEval prompt-level strict 26.1 with thinking off and 17.7 with it on; Falcon-H1-Tiny-90M-Instruct scores 60.8.
• Loops: 8.3% of greedy replies (thinking off) run to the 1,024-token cap in a repeating pattern; 1.2% with the shipped sampling settings.
• Long context: recall fails beyond 16K tokens.

About the evaluation and this dataset:

• The judge is a model. Chat-quality verdicts come from DeepSeek-V4.1-Flash, not from people. Position bias and length bias are measured (judge/.../results.json) but not removed; the MobileLLM-R1.5 comparison is length-confounded.
• Small differences are noise. bf16 greedy decoding is not batch-invariant at this size: for Laghu, 32 of 40 check items gave identical first 512 tokens batched and one at a time (tables/TABLES_full.md, scorer checks).
• Two held-out bits-per-byte sets favour our models. The web and reasoning sets are slices of our own data lineage; the new-text set is the fair comparison.
• Not every model has per-item files. MobileLLM-R1.5-140M and MobileLLM-R1-140M-base appear as aggregate scores and judge verdict counts only, and nltower-100m-base as aggregate scores only, because of their licences (see "Per-model notices"). Their numbers in the tables cannot be re-derived from this dataset.
• Some text is held back. MATH-500, HellaSwag, PIQA, ARC, MMLU and WinoGrande ship as ids; the join script needs network access to the pinned public revisions. Where a model copied MATH-500 problem text into its reply, that text is replaced by a marker (24,674 spans in 2,072 replies) that the join script fills back in.
• A few outputs are redacted. Ten email addresses and phone numbers that models wrote into 8 replies are replaced by [REDACTED-EMAIL] / [REDACTED-PHONE] (rows carry pii_redacted). Scores were computed on the original text.
• One snapshot. Rival models were pinned to the revisions in the manifests on 19 October 2026; newer revisions may score differently.

What is inside

Path	What it holds
chat_suite/prompts/	chat_prompts.jsonl (324, with source, source id and licence), multiturn.jsonl (214), gsm8k_test.jsonl (1,319), ifeval.jsonl (541), math500_l123_ids.jsonl (238 ids, levels and subjects), ITEMS.json (pinned revisions and file hashes)
chat_suite/generations/<model>/<engine>/<task>__<mode>__<variant>.jsonl.gz	one row per item: id, prompt token count and hash, cap, tokens generated, finish reason, raw (the full output, thinking included), period (loop check). meta.json: template, stop tokens, versions. Engines other than the model's main one are the scorer checks (one-at-a-time and fp32 reruns)
chat_suite/scores/<model>/<task>__<mode>[__<variant>].jsonl	per-item scores (GSM8K / MATH ok, IFEval per-instruction booleans, multi-turn copy / right, E2 one / two, loop loop); summary.json (every aggregate, MobileLLM included), validation.json (scorer checks)
chat_suite/judge/deepseek-v4.1-flash/	per-pair verdicts vs SmolLM2-135M-Instruct, Falcon-H1-Tiny-90M-Instruct and Gargi (both orders, reply lengths, the judge's one-sentence reason); results.json (win rates, intervals, p-values, position bias); MobileLLM-R1.5: verdict counts only
chat_suite/manifest.json	every chat model: repo, pinned commit, sampling, context, licence tag, notice
base_suite/<model>/<task>.jsonl.gz	lm-eval 0.4.13 rows: doc_id, gold index, per-choice log-likelihoods, acc / acc_norm, lm-eval's doc_hash / prompt_hash / target_hash; GSM8K rows keep question, few-shot prompt and the model's generation. results.json: lm-eval's aggregate output
base_suite/summary.json	headline numbers for all 27 models (aggregate-only models included), repo, revision, licence
bpb/<set>/<model>.jsonl	per-document bits, bytes and tokens; fresh-text rows carry doc_id (joins to udgamlabs/fresh-text-2026-10) and in_public_set
long_context/	gate_base_suite/ (Laghu-Base and Laghu-Base-32K on validation splits, per item), gate.json, long_eval.json (needle, copy, bits per byte by position), RESULTS_32K.md
tables/	the published tables and method notes: TABLES_full.md, CLAIMS_2B.md, BASE_RESULTS_CUDA.md, CHAT_SUITE_METHODS.md, CHAT_SUITE_NOTES.md, CHAT_SUITE_NOT_RUN.md
join_items.py	re-attaches held-back item text from pinned revisions and checks it
SCAN_REPORT.md, MANIFEST.json	what was checked before release (counts only), and every file with its size and sha256

Model names: laghu is the released chat model (evaluated from udgamlabs/Laghu-SFT-v1 at revision 419343f6a2, the weights released as udgamlabs/Laghu); laghu-base is Laghu-Base (EMA weights at step 76,294, before the 32K stretch); laghu-base-32k is Laghu-Base-32K. "off" / "on" is Laghu's thinking switch; "native" means the model has no switch.

How it was made

• Pre-registration. The chat-suite rules (items, caps, scoring, comparisons, Holm correction, release gate) were written before any Laghu chat result existed; the judge was changed to DeepSeek-V4.1-Flash before measuring, because the pre-registered judge could not be reached. The gate failed (no category won against all three rivals); the release went ahead as a fully open research release, with no result changed.
• Hardware and software. One NVIDIA RTX PRO 6000 Blackwell (96 GB), bf16 for every model; vLLM 0.31.0 for most chat models and Hugging Face transformers 5.19.0 where vLLM lacks the architecture; lm-eval 0.4.13 for the base suite. Every model is pinned to a commit (chat_suite/manifest.json, base_suite/manifest.json, base_suite/remote_code_pins.json). Full method notes: tables/CHAT_SUITE_METHODS.md.
• Scoring. GSM8K: the number after the last "Final answer", else the last number. MATH-500: Math-Verify 0.9.0. IFEval: lm-eval 0.4.13's own scorer, seeded so re-scoring is deterministic. Multi-turn: near-copy and expected-answer checks. Loops: the reply hits its cap and its last 256 tokens repeat. The scorers are in the code repository.
• Checks before release (SCAN_REPORT.md): the chat suite was re-scored from the published generations with the original scorer, and the tables re-rendered from that; every per-item score and every table line matched. Base-suite, bits-per-byte and 32K numbers were recomputed from the per-item files and match the published tables. The release was scanned for secrets (none) and personal data (10 hits in model outputs, redacted; none in the real-user prompts).

Benchmark items

Benchmark	Source and pinned revision	Licence	In this dataset
GSM8K test (1,319)	openai/gsm8k @ 740312ad	MIT	text, ids, outputs, scores
IFEval (541)	google/IFEval @ 966cd895	Apache-2.0	text, ids, outputs, scores
MATH-500 levels 1-3 (238)	HuggingFaceH4/MATH-500 @ 6e4ed1a2	from the MATH dataset, whose original release is disabled on Hugging Face after a takedown notice	ids, outputs (problem text replaced by markers), scores
ARC-Easy, ARC-Challenge	allenai/ai2_arc @ 210d026f	CC-BY-SA-4.0	ids and scores (keeps this dataset under CC-BY-4.0)
HellaSwag	Rowan/hellaswag @ 218ec52e	see the source	ids and scores
PIQA	baber/piqa @ 142f6d73 (lm-eval's copy of ybisk/piqa)	not stated by the source	ids and scores
WinoGrande (winogrande_xl)	allenai/winogrande @ 01e74176	Apache-2.0	ids and scores
MMLU	cais/mmlu @ c30699e8	MIT	ids and scores
Multi-turn follow-ups (214)	written for these evaluations by Udgam Labs with Claude (Anthropic)	CC-BY-4.0	text, outputs, scores
Chat prompts (324)	real-user prompts: WildChat-4.8M (147, ODC-By), OpenAssistant oasst2 (104, Apache-2.0), LMArena arena-human-preference-140k (66) and -100k (7) (user prompts CC-BY-4.0)	as the source	text with source ids, outputs, judge verdicts

The 324 chat prompts keep their source ids so that removal requests to the source datasets (for example WildChat's) can be honoured here too; to ask for a prompt's removal, open a discussion on this dataset's page. Publishing them ends their use as a private held-out set.

Per-model notices

Model outputs in this dataset are machine-generated and may be wrong. Judge outputs are machine-generated by DeepSeek-V4.1-Flash through the DeepSeek API and may contain errors; naming the judge implies no endorsement.

• Falcon-H1-Tiny (Falcon-H1-Tiny-90M-Instruct, -R-90M, -Multilingual-100M-Instruct, -90M-Base): made by the Technology Innovation Institute (TII), licensed under the TII Falcon License (December 2024) and the Falcon Acceptable Use Policy, https://falconllm.tii.ae/falcon-terms-and-conditions.html. TII claims no rights in outputs; please use these outputs in line with the Acceptable Use Policy.
• Gemma 3 (gemma-3-270m-it, gemma-3-270m): Gemma is provided under and subject to the Gemma Terms of Use, https://ai.google.dev/gemma/terms, and the Gemma Prohibited Use Policy. Google claims no rights in outputs. Warning: under the Gemma Terms, a model trained on Gemma outputs so that it performs like Gemma may be a Model Derivative of Gemma, and the Gemma Terms then apply to it.
• LFM2.5-230M (Liquid AI): licensed under the LFM Open License v1.0.
• BananaMind-2-Pro: licensed under the BananaMind Community License 1.0.
• MobileLLM (MobileLLM-R1.5-140M, MobileLLM-R1-140M-base): MobileLLM results are reported for noncommercial research under the FAIR Noncommercial Research License; we acknowledge Meta's materials. Aggregate scores and judge verdict counts only; no MobileLLM output text is in this dataset.
• nltower-100m-base: the model has no licence; aggregate scores only.
• All other evaluated models are Apache-2.0 (Blaze-SFT: MIT); see chat_suite/manifest.json and base_suite/summary.json for each model's licence tag.

Intended use and limitations

For research on small language models: checking our numbers, re-scoring with other rules, error analysis, and comparing new models on the same items. The outputs are not a training set. They include wrong answers and looping text by design, and training a model on some of them can bring the source model's terms with it (see the Gemma warning above). Do not use the model outputs as facts.

Licence

• Our outputs, scores, verdict tables and annotations (everything Udgam Labs produced: Laghu's, Gargi's and Udgam-384M's outputs, every per-item score, the multi-turn items, file layouts and summaries): CC-BY-4.0, © 2026 Udgam Labs LLP.
• Other models' outputs are released under CC-BY-4.0 as far as we hold rights in them, together with the notices above.
• Prompts and benchmark items keep their upstream licences (table above). WildChat prompts: ODC-By 1.0; oasst2: Apache-2.0; LMArena user prompts: CC-BY-4.0; GSM8K: MIT; IFEval: Apache-2.0.
• MobileLLM aggregates are reported for noncommercial research under the FAIR Noncommercial Research License.

Citation

@misc{udgamlabs2026laghuevals,
  title        = {Laghu evaluation outputs},
  author       = {{Udgam Labs}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/udgamlabs/laghu-evals}}
}

Laghu 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