Update README.md

README.md

@@ -194,9 +194,11 @@ Note that any examples whose inputs did not fit within the token limit were left
 
 ### SteamSHP - An Open-Source Preference Model
 
-We have finetuned a 6 billion-parameter FLAN-T5-XL model called [SteamSHP](https://huggingface.co/stanfordnlp/SteamSHP-preference-model) on both the SHP dataset and the helpfulness data from Anthropic's HH-RLHF.
-It achieves 72.8% across all domains (including 73.1% on the HH-RLHF alone).
-We encourage you to use it for NLG evaluation, for building reward models for RLHF, or for another purpose you deem fit!
+We have finetuned two FLAN-T5 models on both the SHP dataset and the helpfulness data from Anthropic's HH-RLHF. They are
+- [SteamSHP-XL](https://huggingface.co/stanfordnlp/SteamSHP-flan-t5-xl), a 3B parameter model that achieves 72.8% on the test data.
+- [SteamSHP-Large](https://huggingface.co/stanfordnlp/SteamSHP-flan-t5-large), a 780M parameter model that achieves 72.0% on the test data.
+
+We encourage you to use SteamSHP for NLG evaluation, for building reward models for RLHF, or for another purpose you deem fit!
 
 
 ## Biases and Limitations