End of training

README.md

@@ -12,25 +12,37 @@ should probably proofread and complete it, then remove this comment. -->
 
 # ProGemma
 
-This is a custom configuration of Google's Gemma 2 model that is being pre-trained on amino acid sequences of lengths 0 to 512. I used the free version of Google Colab to train this model, so updates are made regularly as the model hits new checkpoints. As of 08.08.2024, the model has been trained on about 25% of the dataset.
+This model is a fine-tuned version of [JuIm/ProGemma](https://huggingface.co/JuIm/ProGemma) on an unknown dataset.
 
-The model generates amino acids on a letter-by-letter basis.
+## Model description
 
-Current training loss is about 2.7. Preliminary evaluation of generated sequences on AlphaFold 3 shows pTM scores of ~0.4 and average pLLDT scores ~60. After training is complete, a proper evaluation will be done to see whether sequences result in proteins with a low free energy. Perplexity scores are on par with NVIDIA's ProtGPT2 for a generated sequence of a given length. 
+More information needed
 
-The purpose of this model was to see whether I could develop an alternative to NVIDIA's ProtGPT2. ProGemma also serves as a stepping stone to a new model that will also utilize control tags to generate proteins based on function.
+## Intended uses & limitations
 
-To use this mode for yourself using the pipeline within the Transformers package, please see the code below:
+More information needed
 
-from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
+## Training and evaluation data
 
-model = AutoModelForCausalLM.from_pretrained("JuIm/ProGemma") tokenizer = AutoTokenizer.from_pretrained("JuIm/Amino-Acid-Sequence-Tokenizer")
+More information needed
 
-progemma = pipeline("text-generation", model=model, tokenizer=tokenizer)
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 0.001
+- train_batch_size: 1
+- eval_batch_size: 8
+- seed: 42
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_ratio: 0.4
+- training_steps: 7000
+
+### Training results
 
-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)
 
-print(sequence)
 
 ### Framework versions
 

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:91d2d4fd141b20914c5e884128cdb1caf7fa4d935ed311d6c2b4d285cb3663b2
+oid sha256:63ca2e32f9fb32e0813527bea7dbb58b2950f73ca8208c66e54952e79a7da3db
 size 1101271208

runs/Aug09_14-42-52_7ac9000e2e30/events.out.tfevents.1723214576.7ac9000e2e30.414.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25e6883ee38aabbb6278224eabf42282164cf7725fd82a04edea8934793b417b
+size 1481827

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:575b5f0f90417367e8f856bc1811c0f7ac60cdb30e364b53d2f56fa6476341b5
+oid sha256:a55f5773af82c07ba51762012c6521e9198e116cd1f0a20f0fed73d39a45f94d
 size 5112