Describe training by time, not cost

app.py

@@ -97,8 +97,8 @@ def chat_fn(message, history, max_tokens, temperature):
 
 ABOUT = """
 **A 20-layer chat model (~390M transformer + ~500M embedding parameters) trained completely from scratch** — tokenizer, pretraining and chat fine-tuning,
-no pretrained weights anywhere — using [karpathy/nanochat](https://github.com/karpathy/nanochat) on a single rented
-H100 for about **$13** (3.5B tokens, ~6 h).
+no pretrained weights anywhere — using [karpathy/nanochat](https://github.com/karpathy/nanochat) on a single H100:
+**~5 hours of pretraining** on 3.5B tokens plus **47 minutes of chat fine-tuning**.
 
 It writes fluent English, holds a conversation, writes poems and explains things simply, but it **confidently makes up
 facts** (it is ~200x smaller than GPT-3.5). This free CPU Space is slow: replies stream at a few words per second.
@@ -108,7 +108,7 @@ Weights: [sumitkClasses/nanochat-d20-chat](https://huggingface.co/sumitkClasses/
 demo = gr.ChatInterface(
     chat_fn,
     type="messages",
-    title="nanochat-d20 · a chat model trained from scratch for $13",
+    title="nanochat-d20 · a chat model trained from scratch",
     description=ABOUT,
     additional_inputs=[
         gr.Slider(32, 1024, value=min(DEFAULT_MAX_TOKENS, 1024), step=32, label="Max new tokens"),