Update README.md

README.md

@@ -8,9 +8,9 @@ pipeline_tag: question-answering
 ---
 # cosmosage
 
-Cosmosage is a natural-language cosmology assistant that you can ask questions about cosmology.
+Cosmosage is a natural-language cosmology assistant that can answer questions about cosmology.
 
-Fine tune of Mistral-7B-v0.1 on cosmology datasets. Intended for Q&A mode, where the model is given a single question and it replies with a single answer. 
+cosmosage_v0.2 is a fine tune of Mistral-7B-v0.1 on various cosmology-related datasets including open-access textbooks and scientific publications. It is intended to be used in Q&A mode, where the model gives a single answer in response to a single question. 
 
 See https://github.com/tijmen/cosmosage for more details.
 
@@ -52,7 +52,7 @@ ASSISTANT:
 
 ## Qualitative evaluation
 
-cosmosage_v0.2 performs much better than cosmosage_v0.1. While v0.1 did not seem to have picked up much knowledge from the ArXiV papers it was trained on, v0.2 can often give surprisingly good answers to highly technical questions about cosmology. It gives certain answers which it could not have known without having read these recent papers, leading me to conclude that it has learned knowledge from the ArXiV papers. 
+cosmosage_v0.2 performs much better than cosmosage_v0.1. While v0.1 did not seem to have picked up much knowledge from the ArXiV papers it was trained on, v0.2 can give surprisingly good answers to highly technical questions about cosmology. It gives certain answers which it could not have known without having read these recent papers, leading me to conclude that it has picked up some knowledge from the ArXiV papers. 
 
 I've also been impressed by cosmosage's knowledge about astronomy, as well as other branches of physics. However, in these areas it is less clear how much the performance is due to the pretraining of the Mistral model versus the fine-tuning I did.