Introducing SLM-Bench: A Benchmark Built for Tiny Models
#1
by PY-AI-Dev - opened
SLM-Bench: A Benchmark Built for Tiny Models
When you train a 10-million-parameter language model, the standard benchmarks don't tell you much. ARC, HellaSwag, MMLU — these were designed for models with billions of parameters. Asking a 10M model to compete on them is like measuring a bicycle's top speed on a Formula 1 track.
That's why we built SLM-Bench: a benchmark calibrated for sub-10M models, covering six capability areas with 3,000 questions — and it's 100% programmatically generated. No manual annotation, no LLM-generated questions, no human bias.
What SLM-Bench Measures
| Category | What It Tests | Example |
|---|---|---|
| Arithmetic | Basic computation with plausible wrong answers | What is (43 + 5) * 5 - 1? → 239 |
| Pattern | Sequence reasoning and rule detection | What comes next: 2, 6, 18, 54, ? → 162 |
| Grammar | Syntactic understanding | She ___ to school every day. → goes |
| Vocabulary | Word relationships and definitions | Antonym of abundant → scarce |
| Logic | Syllogistic reasoning | All dogs are mammals. All mammals are animals. → All dogs are animals? → True |
| Word Analogy | Relational reasoning | hot : cold :: up : ? → down |
Fully Programmatic: Zero Manual Effort
Every single question is generated by code. The generation logic is open-source and reproducible:
git clone https://github.com/Liodon-AI/slm-bench
cd slm-bench
pip install nltk numpy datasets huggingface_hub
python assemble.py --n_per_category 500 --seed 42
Using SLM-Bench
With lm-evaluation-harness
lm_eval --model hf \
--model_args pretrained=liodon-ai/slm-10m,trust_remote_code=True \
--tasks slm_bench \
--device cuda:0 --batch_size 64
With HuggingFace Datasets
from datasets import load_dataset
ds = load_dataset("liodon-ai/slm-bench", "arithmetic")
print(ds["test"][0])
Why This Matters
- Reproducibility: The generation code IS the dataset specification
- Scalability: Want 10K questions? Change one parameter
- Transparency: Every question traces back to its generation logic
- Cost: $0 in annotation costs
What's Next
- Difficulty calibration based on model performance
- New categories: code comprehension, reading comprehension, factual recall
- Cross-lingual extensions
- Longitudinal tracking of capability emergence during training
Links:
Released under Apache-2.0 by Liodon AI