Prabhāsa-b_ss 0.2 — BabyLM-2026 Strict-Small (10M)
ELC encoder, pure-MLM + RoPE + N-hot + structured masking, Muon optimizer (scale-dependent: Muon wins at 10M).
Official BabyLM-2026 scorer: BLiMP 62.74 (supplement, EWoK, COMPS in repo).
- +3.3 pp over v0.1 (
qbz506/prabhasa-b_ss-0.1, 59.46); −2.3 pp from the GPT-2 baseline (65.08). BabyLM-compliant (≤10 epochs, 10M budget).
