Update README.md

README.md

@@ -12,6 +12,26 @@ should probably proofread and complete it, then remove this comment. -->
 
 # ProGemma
 
+This is a custom configuration of Google’s Gemma 2 LLM that is being pre-trained on amino acid sequences of 512 AA or less in length. Periodic updates are made to this page as training reaches new checkpoints. 
+
+The purpose of this model was to investigate the differences between ProGemma and ProtGPT (GPT-2 architecture) as it pertains to sequence generation. 
+As of 8.22.2024, ProGemma has been trained on ~80% of the training dataset and is still on epoch 1. Training loss is ~2.6. Perplexity scores as well as AlphaFold 3’s ptm, pLDDT, and iptm scores are generally in line with ProtGPT’s scores for sequence lengths < 250, although the testing phase is still very early. I have yet to do testing for sequence lengths > 250. More robust testing is also required for lengths < 250 AA. In my very preliminary testing, HHblit scores of ~0.1 are achieved with relatively easily.
+
+Controlled generation is not a capability of this model, and therefore serves as a method to significantly improve generation as, in principal, a sequence that performs a given function or resides in a particular cellular location can be generated. 
+
+In sequence generation, a top_k of 950 appears to work well as it prevents repetition. This is also seen in ProtGPT.
+
+Below is code using the Transformers library to generate sequences using ProGemma.
+
+
+from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
+model = AutoModelForCausalLM.from_pretrained("JuIm/ProGemma")
+tokenizer = AutoTokenizer.from_pretrained("JuIm/Amino-Acid-Sequence-Tokenizer")
+progemma = pipeline("text-generation", model=model, tokenizer=tokenizer)
+sequence = progemma("<bos>", top_k=950, max_length=100, num_return_sequences=1, do_sample=True, repetition_penalty=1.2, eos_token_id=21, pad_token_id=22, bos_token_id=20)
+s = sequence[0]['generated_text']
+print(s)
+
 
 ### Framework versions