Update CyberSLM-33M-Base / -Instruct with retrained checkpoints

#92

Update CyberSLM-33M-Base / -Instruct (retrained)

This updates the two existing CyberSLM rows rather than adding new ones: the
models have been retrained from scratch after fixing 19 defects in the
training pipeline, and these rows now point at the retrained checkpoints.

Pretrained from scratch on a 194.8M-token corpus that is ~60%
cybersecurity, ~20% general English, ~15% programming, ~5% CS fundamentals.
Open weights, training checkpoints, not merged.

Setup

harness lm-evaluation-harness 0.4.9 (HellaSwag, ARC-E, ARC-C, PIQA)
arithmark AxiomicLabs' bencharithmark-3.py and benchmark_arithmark-2.0.py, unmodified
n-shot 0
dtype float32
stack torch 2.13.0+cu130, transformers 4.57.6, datasets 5.0.1
hardware NVIDIA A10G
architecture 12L, d_model 384, 6 heads, SwiGLU 1024, RoPE, RMSNorm, tied head, ctx 2048, vocab 32,000

Results

Both acc and acc_norm are given, because the column convention is not
documented. The headline row uses the mapping I inferred by calibrating
against the SmolLM2-135M row already on the board (piqa tracks acc;
hellaswag and arcChall track acc_norm; arc tracks arc_easy acc). If that
mapping is wrong, please use whichever column is correct from the table below.

model metric ARC-E HellaSwag PIQA ARC-C ArithMark-3 ArithMark-2
CyberSLM-base acc 32.79 26.35 54.68 18.17 31.00 26.96
CyberSLM-base acc_norm 32.24 26.84 52.23 22.70 31.00 -
CyberSLM-instruct acc 32.37 26.51 53.81 19.11 30.00 26.88
CyberSLM-instruct acc_norm 32.15 26.19 52.88 22.35 30.20 -

Honest note on these numbers

These scores sit at or near chance on HellaSwag, PIQA and ARC-Challenge, and
are comparable to models an order of magnitude smaller. That is expected and I
would rather state it than have it discovered: the corpus is domain-specific,
and these benchmarks measure everyday commonsense, physical reasoning and
grade-school science, none of which the training data covers.

The pipeline itself is verified, so the numbers are not an artefact of a broken
setup: the HF export reproduces the original implementation's logits to 3.1e-05
with identical greedy decoding, and the tokenizer matches SentencePiece exactly
on all 9,000 ArithMark-3 strings. Everything is reproducible from the repo.

Happy to re-run under different settings if you use a different n-shot or
metric convention.

sabari2005 changed pull request title from Add CyberSLM-base, CyberSLM-instruct to Update CyberSLM-33M-Base / -Instruct with retrained checkpoints
Datdanboi25 changed pull request status to merged
Axiomic Labs org

Hi Sabari, cool models, and thank you for your submission, however all benchmarks should be acc norm!

Sign up or log in to comment