Sparrow-1M iter6a baseline (3-digit multiplication)

The strongest verified Sparrow-1M baseline as of 2026-05-07.

  • Task: 3-digit multiplication, in-distribution
  • Accuracy: 76% on held-out eval (1.3M params)
  • Format: byte-level, no BPE, per-digit space-separated, MSB-first operands and result
  • Architecture: Qwen3 dense, 5 layers, hidden 128, GQA 4Q-2KV, RoPE base=10000
  • Training: 25K steps, batch 64, peak_lr 1.5e-3, warmup 500, cosine to 1e-5, 1M procedurally-generated samples
  • Author: Crownelius (github.com/Crownelius)
  • Code: github.com/Crownelius/crowfeather-50m-v1 (sparrow_1m subfolder, commit 45a9ea6 era)

This baseline is the comparison point for iter37's Position Coupling negative result and for the planned iter38a/b/c Abacus-style ablations on resume.

Downloads last month
5
Safetensors
Model size
1.26M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support