Skip to content
D-CSIL

Model Registry · D-CSIL family

SLM-10M

Causal transformer · 12 layers · 256 dim · GQA 8Q/2KV · QK-norm · RoPE θ=100k · SwiGLU · weight tying

SLM-10M — TRAINING LOSS (DEMO DATA)48,828 STEPS · MIN 3.60
3.64.45.36.16.90k12k24k37k49k

Benchmarks

MetricValueNotes
HellaSwag (acc_norm)26.53%
ARC-Easy (acc_norm)30.47%
ARC-Challenge (acc_norm)25.00%
PIQA (acc_norm)50.92%
ArithMark-2.024.32%
Leaderboard average32.38%zero-shot, Open SLM harness

Known Limitations

  • Open-ended generation shows repetition loops — evaluated via log-likelihood ranking, as appropriate at 10M.
  • Loss-curve anchors are approximate; benchmark scores are exact from the 2026-06-23 eval files.