TobiasLogic/barunlm-35m-verification

BarunLM-35M benchmark verification

Independent reproduction of the evaluation claims in harrrshall/BarunLM-35M. Full writeup in the model's discussion thread.

Method

Same 9 tasks as the original card, 0-shot, lm-eval-harness==0.4.12, seed 1234, on the exact revisions pinned in the model's own benchmark_results.json:

• ARC-Challenge, ARC-Easy, BoolQ, HellaSwag, LAMBADA OpenAI, OpenBookQA, PIQA, SciQ, WinoGrande

No decontamination filter was applied (BarunLM's card excludes 1,854 samples that overlap its training corpus; that scan is specific to their training data and not reproducible from outside, so these numbers are on the full standard test/val splits).

model.safetensors SHA256 and parameter count (35,072,768) were confirmed to match the published card.

Files

• barunlm_eval.py: custom lm-eval-harness model wrapper for BarunLM's architecture (it ships its own PyTorch classes, not a standard AutoModelForCausalLM). Includes two patches for bugs hit while running it on CUDA/bf16, both documented inline: a torch.compile(flex_attention) path that failed to compile on the test GPU (worked around by forcing the model's own eager SDPA fallback), and a dtype bug where the local-attention causal mask is never cast to the query's dtype, which crashes under bf16 specifically. To use, drop this file alongside a checkout of the barunlm/ package and weights from the original repo.
• run_eval.py: CLI runner used for every model in this comparison.
• remote_run_all.sh / remote_run_resume.sh: orchestration scripts used on the rented GPU boxes.
• results/: raw lm-eval-harness output JSON for all 6 models (barunlm-35m, lfm25-230m-base, pythia-160m-deduped, stentor-30m, tinystories-33m, pythia-70m-deduped), plus lfm25-230m-base-recheck-gpu2.json, a second independent run of LFM2.5-230M-Base on a separate rented GPU (see below).
• summary.json: computed macro accuracy and per-task breakdown, reproduced vs. claimed.

Results

Macro accuracy is the unweighted mean of acc,none across the 9 tasks, matching the metric BarunLM's own per-task numbers correspond to.

Model	Reproduced	Claimed
BarunLM-35M	41.37%	41.01%
LFM2.5-230M-Base	49.57%	39.20%
Pythia-160M-deduped	35.47%	37.35%
Stentor-30M	35.64%	36.46%
TinyStories-33M	31.74%	33.16%
Pythia-70M-deduped	29.00%	31.71%

BarunLM-35M's own number and its margins over the four smaller peer models reproduce closely (within the range the missing decontamination filter would explain). LFM2.5-230M-Base does not: it scores 10.4 points higher than claimed, meaning it actually outperforms BarunLM-35M by about 8 points rather than losing to it.

LFM2.5-230M-Base reproducibility check

Because that result runs counter to the model card's central claim, it was re-run on a second, independent rented GPU (different physical machine, fresh download, fresh environment, identical protocol). All 9 per-task scores matched the first run exactly, macro accuracy 49.57% both times. See results/lfm25-230m-base-recheck-gpu2.json and the lfm25_230m_base_reproducibility_check block in summary.json.