Update README.md

README.md

@@ -68,7 +68,8 @@ Per un'azienda: **un solo Genome enterprise** + Soul intercambiabili per ogni re
 | **V1** | 15,000 | 1.07 | 1.5GB curated | Baseline — `zeta25b_step15000.pt` |
 | **V2 Repair** | 500 | 1.07 | 1.4GB SFT mix | Fine-tuning da V1 |
 | **V3 Knowledge** | 5,000 | 1.33 | **9.1GB** (Wiki EN/IT, C4, Books) | Espansione conoscenza |
-| **V4 Expanded** | 10,000 | In corso | 9.1GB | LoRA rank 512 (~950M trainabili) |
+| **V4 Expanded** | 10,000 | **1.28** | 9.1GB | LoRA rank 512 (~950M trainabili) ✅ |
+| **V5 Code** | 5,000 | In training | 4.5GB code | Code Specialist Soul 🔄 |
 
 - Hardware: **NVIDIA A40 48GB** (singola GPU)
 - Architettura: Fractal Gated Causal TCN, No-Attention
@@ -135,8 +136,8 @@ DOI: `10.5281/zenodo.18622610`
 - [x] V1: Genome + Soul baseline (15K step, loss 1.07)
 - [x] V2: Fine-tuning repair
 - [x] V3: Knowledge expansion (9.1GB dataset, Wiki EN/IT, C4, Books)
-- [/] V4: LoRA rank 512 (~950M trainabili) — **in training ora**
-- [ ] V4 Code-Only: Fork specializzato codice
+- [x] V4: LoRA rank 512 (~950M trainabili) — **loss 1.28, PPL 3.6** ✅
+- [/] V5 Code: Code Specialist Soul (4.5GB code dataset) — **in training ora**
 - [ ] GGUF v2: Conversione e upload del checkpoint V4
 - [ ] Scaling 50B: Stesso Genome, Soul espansa a 64 layer
 - [ ] Scaling 1T: Dimostrazione proof-of-concept