Update README.md

README.md

@@ -21,10 +21,12 @@ datasets:
 
 ## Fast Anime PromptGen
 
-Trained on the 80K Safebooru prompts
+`pytorch_model` is trained on 80K anime tags, all with `up_score` greater than 8 and without "greyscale","girls","boys", and "others", fetched from the [Safebooru API](https://safebooru.donmai.us/posts/random.json). 
+I didn't release the V1 model because it only generated gibberish prompts. After trying all means to correct that behavior, I eventually figured that the cause of the gibberish prompts is not from the model or training duration, but rather from the random usernames present in the training data. 
+Here's the complete [prompt preprocessing](./preprocess.py).
+
 
 Todo:
-- complete data preprocessing and training description
 - upload Danbooru model
 
 ## Text-to-image Examples
@@ -48,7 +50,7 @@ tokenizer = GPT2Tokenizer.from_pretrained('distilgpt2')
 tokenizer.add_special_tokens({'pad_token': '[PAD]'})
 model = GPT2LMHeadModel.from_pretrained('FredZhang7/anime-anything-promptgen')
 
-prompt = r'1girl, genshin impact'
+prompt = r'1girl, genshin'
 
 # generate text using fine-tuned model
 nlp = pipeline('text-generation', model=model, tokenizer=tokenizer)