fromziro/ZeroS-Qana-5M

Update repo links and citation URLs to fromziro/ZeroS-Qana-5M

Update repo id in training config

Update model card for ZeroS-Qana-5M

Update README.md

Preserve historical CPU RoPE cache values

Fix config-driven RoPE cache initialization

Move RoPE parameters into model config (#2)

Update modeling_ac_swiglu.py

Update modeling_ac_swiglu.py

Rebrand model as Qana-mini-5M

Add BLiMP and standalone ArithMark-3 results

Report transparent task-level evaluation results

List authors individually in attribution

Clarify lead authorship and open attribution terms

Open full model and architecture under MIT

Document credits and scoped model licensing

Explain the ChannelMix-SwiGLU mechanism

Rebrand as ChannelMix-SwiGLU Mini and refresh model card

Use official Open SLM leaderboard scoring

Step 40000: Int 7.55, Avg 36.36%, BPB 1.4241

Step 30000: Int 7.35, Avg 36.19%, BPB 1.4807

Step 20000: Int 5.94, Avg 35.49%, BPB 1.4817

Step 10000: Int 5.81, Avg 35.46%, BPB 1.5028

Step 1000: Int 4.59, Avg 34.55%, BPB 1.6713

initial commit