Update README.md

README.md

@@ -163,7 +163,7 @@ __*Only train splits are used, and a decontamination by cosine similarity is per
 
 ## Prompt formatting
 
-In sticking with the theme of the bagel, I didn't want to use a single prompt format, so I used 4 - vicuna, llama-2, alpaca, and chat-ml.
+In sticking with the theme of the bagel, I didn't want to use a single prompt format, so I used 4 - vicuna, llama-2, alpaca, and a modified chat-ml.
 I also didn't want to randomly select a single prompt format for each item (hoping each instruction would generalize more when used in a variety of prompt formats), so each instruction is converted into every prompt format (with 0.75 probability).
 
 This means each epoch of our fine-tune is the equivalent of 3 epochs.
@@ -223,12 +223,14 @@ print(tokenizer.apply_chat_template(chat, tokenize=False))
 </details>
 
 <details>
-  <summary><b>ChatML</b></summary>
+  <summary><b>ChatML (sort of)</b></summary>
+
+  ChatML special tokens are really obnoxious, so instead of enlarging the tokenizer and embedding layers (which decreases performance and causes inference problems in tensor parallelism), I just use BOS and EOS tokens instead of `<|im_start|>` and `<|im_end|>` - and no, I won't change this.
 
   ```text
-  {bos}<|im_start|>{role}
+  {bos}{role}
   {text}
-  <|im_end|>{eos}
+  {eos}
   ```
 </details>