Update README.md

README.md

@@ -9,23 +9,23 @@ license: apache-2.0
 # Model Card for Ostrich
 
 
-- Contentious, judgemental, uncensored, can't agree with itself 32% of the time!
-- Trained a bit about nostr
+- Trained with some of the Nostr notes
 - Trained a bit about bitcoin
-- Trained a bit in the health domain
-
-I am having success with chat template: \<s\> \[INST\] ... \<\/s\> 
-
-It may also work with ChatML format, though I see more repetitions when I use that.
+- Aligned a bit in these domains:
+- Health
+- Permaculture
+- Phytochemicals
+- Alternative medicine
+- Herbs
+- Nutrition
 
+Read more about it here:
+https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzde3xsunjwfkxcunwv3jvtnjyc
 
 ## Model Details
 
-Based on https://huggingface.co/crestf411/daybreak-miqu-1-70b-v1.0-hf because it is one of the most uncensored according to https://huggingface.co/spaces/DontPlanToEnd/UGI-Leaderboard.
-
 
-- **Fine tuned by:** someone
-- **Finetuned from model:** https://huggingface.co/crestf411/daybreak-miqu-1-70b-v1.0-hf
+- **Finetuned from model:** https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct
 
 
 ## Uses
@@ -34,6 +34,10 @@ Ask any question, compared to other models this may know more about Nostr and Bi
 You can use llama.cpp to chat with it.
 You can also use llama-cpp-python package to use it in a Python script.
 
+Llama3 chat template can be used. <|begin_of_text|><|start_header_id|> ...
+
+Use repeat penalty of 1.05 or more to avoid repetitions.
+
 
 ## Warning
 
@@ -45,10 +49,11 @@ The trainer, developer or uploader of this model does not assume any liability.
 
 ### Training Data
 
-Nostr related info from web and nostr itself, bitcoin related info.
+Nostr related info from web and nostr itself, bitcoin related info. Info on health domain.
+Information that aligns well with humanity is preferred.
 
 ### Training Procedure
 
 LLaMa-Factory is used to train on 2x3090! fsdp_qlora is the technique. 
 
-It took ~185 hours for a dataset of 122MB.
+The Nostr training took ~30 hours for a dataset of about 20MB.