diff --git a/eval/README.md b/eval/README.md
index f69cb25..f4a076e 100644
--- a/eval/README.md
+++ b/eval/README.md
@@ -6,7 +6,10 @@ benchmark’s evaluation protocol. For AlpacaEval 2, we report both the raw win
 length-controlled win rate (LC). The LC metric is specifically designed to be robust against model verbosity. For Arena-Hard, we report the win rate (WR) against the baseline model. 
 
 ### AlpacaEval 2
-We provides generation configurations for the released models in the `alpacaeval2/configs` directory, and the corresponding generation templates can be found in `alpacaeval2/templates`. To evaluate the models on AlpacaEval 2, please use the [`alpaca-eval`](https://github.com/tatsu-lab/alpaca_eval) package.
+We provide generation configurations for the released models in the `alpacaeval2/configs` directory, and the corresponding generation templates can be found in `alpacaeval2/templates`. To evaluate the models on AlpacaEval 2, please use the [`alpaca-eval`](https://github.com/tatsu-lab/alpaca_eval) package.
 
 ### Arena-Hard
 We provide generation configurations for the released models in the `arenahard/configs` directory, and the corresponding generation templates can be found in `arenahard/templates`. To evaluate the models on Arena-Hard, please use the [`arena-hard-auto`](https://github.com/lm-sys/arena-hard-auto) package.
+
+### MT-Bench
+We provide the reference answers generated by GPT-4 Turbo in the `mt-bench` directory which are more accurate than the original GPT-4 generated answers. To evaluate the models on MT-Bench, please use the [`FastChat LLM Judge`](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge#mt-bench) package and replace the [reference answers generated by GPT-4](https://github.com/lm-sys/FastChat/blob/main/fastchat/llm_judge/data/mt_bench/reference_answer/gpt-4.jsonl).