Update README.md

Update README.md

Add training config (carried over from kotodama-3b-base before dupe deletion)

Reconcile model card with kotodama-3b-base dupe: 32-source data mix (was '17-source'), load snippet, bf16-eval warning, config pointer

Model card with final 4-way eval results

3B base final checkpoint, step 195311 (384B tokens + 19.5k-step cosine cooldown)

initial commit