Update card to 4000-step checkpoint: ppl 268.57→55.50, val loss 3.8398→3.7659

README.md

@@ -38,19 +38,18 @@ A 6.95M-parameter GPT-2 style language model trained **from scratch** on [TinySt
 
 - **Data**: TinyStories (~10M BPE tokens after tokenization)
 - **Batch size**: 32 sequences × 512 tokens
-- **Steps**: 3,175 (best checkpoint)
+- **Steps**: 4,000 (best checkpoint)
 - **LR schedule**: Cosine decay with warmup
 - **Hardware**: 32-core CPU, ~2 hours
-- **Best val loss**: 3.8398
+- **Best val loss**: 3.7659
 
 ## Evaluation
 
 | Metric | Value |
 |--------|-------|
-| Perplexity (held-out TinyStories, 100×512) | 268.57 |
-| Perplexity (mid-dataset, 50×512) | 302.45 |
+| Perplexity (held-out TinyStories, 100×512) | 55.50 |
 
-The held-out perplexity is computed on the last 2M tokens (not seen during training). The gap between train val loss (3.84) and held-out perplexity (268.6) reflects the difficulty of the TinyStories distribution at this model size.
+The held-out perplexity is computed on the last 2M tokens (not seen during training). The gap between train val loss (3.77) and held-out perplexity (55.5) reflects the difficulty of the TinyStories distribution at this model size.
 
 ## Why subword?
 
@@ -76,4 +75,4 @@ print(tok.decode(outputs[0], skip_special_tokens=True))
 - 512-token context window
 - Will produce repetitive or incoherent text on out-of-distribution inputs
 - Not a chat model, not instruction-tuned
-- Quality is limited by the 7M parameter budget
+- Quality is limited by the 7M parameter budget