Switch org-card benchmarks to accuracy metrics (higher is better), run on local GPU

README.md

@@ -16,16 +16,13 @@ pinned: false
 - [LUNA-300M](https://huggingface.co/ASTERIZER/LUNA-300M) — 4.5B-token pretrain (2.7× scale-up)
 - LUNA-1B — in progress
 
-**Scaling benchmarks — 100M → 300M** *(latest base checkpoints, same 4.5B-token corpus, held-out ~100K-token eval: EN / KN / TE / TA / HI + code)*
+**Scaling benchmarks — LUNA-100M → LUNA-300M** *(latest base checkpoints, same 4.5B-token corpus, held-out ~100K-token eval on RTX 4060 Ti — higher is better)*
 
 | metric | LUNA-100M | LUNA-300M | gain |
 |---|---|---|---|
-| Eval loss | 3.88 | 3.81 | ↓1.6% |
-| Perplexity — overall | 48.2 | 45.2 | ↓6.1% |
-| Perplexity — English | 41.9 | 35.2 | ↓16.0% |
-| Perplexity — code | 51.9 | 43.9 | ↓15.4% |
-| Perplexity — Kannada | 15.7 | 13.5 | ↓13.8% |
-| Inference (CPU fp32) | 3,303 tok/s | 1,222 tok/s | 2.7× params |
+| Next-token prediction accuracy | 32.5% | 32.7% | ↑0.8% |
+| Word-prediction accuracy (LAMBADA-style) | 16.3% | 17.1% | ↑4.4% |
+| Throughput (RTX 4060 Ti) | 39.2K tok/s | 14.5K tok/s | 2.7× params |
 
 **Tokenizer (multilingual BPE)**
 - [LUNA-1B-Tokenizer](https://huggingface.co/ASTERIZER/LUNA-1B-Tokenizer) — 128K / 64K / 8K vocab