Upload README.md with huggingface_hub

README.md

@@ -43,7 +43,6 @@ pip install esm@git+https://github.com/Biohub/esm.git@c94ed8d
 ```
 
 ```py
-# pip install esm@git+https://github.com/Biohub/esm.git@c94ed8d
 import torch
 from transformers import AutoModelForMaskedLM, AutoTokenizer
 
@@ -65,7 +64,6 @@ print(f"logits shape: {tuple(output.logits.shape)}")
 By default, the model returns only the final layer representations. To return hidden states from **all transformer layers**, set:
 
 ```py
-# pip install esm@git+https://github.com/Biohub/esm.git@c94ed8d
 output = model(**inputs, output_hidden_states=True)
 ```
 
@@ -147,7 +145,6 @@ Performance metrics are detailed in our [ESMC & ESMFold2 paper](https://biohub.a
 Instead of scaled dot product attention (sdpa) you can use a flash attention backend. This requires running the model in bfloat16.
 
 ```py
-# pip install esm@git+https://github.com/Biohub/esm.git@c94ed8d
 model = (
     AutoModelForMaskedLM.from_pretrained(
         "biohub/ESMC-6B",
@@ -169,7 +166,6 @@ To get interpretable features from ESMC 6B hidden states and per-layer MLP outpu
 * [ESMC SAEs for MLP outputs (all layers)](https://huggingface.co/collections/biohub/esmc-saes-for-mlp-outputs-all-layers)
 
 ```py
-# pip install esm@git+https://github.com/Biohub/esm.git@c94ed8d
 import torch
 from transformers import AutoModel, AutoTokenizer
 
@@ -202,7 +198,6 @@ for i, sae_out in enumerate(output.sae_outputs):
 ESMC can predict masked amino acids and compute the corresponding loss:
 
 ```py
-# pip install esm@git+https://github.com/Biohub/esm.git@c94ed8d
 import torch
 from transformers import AutoModelForMaskedLM, AutoTokenizer
 
@@ -229,7 +224,6 @@ print(f"Loss: {output.loss.item():.6f}")
 ### Fine-tuning with peft
 
 ```py
-# pip install esm@git+https://github.com/Biohub/esm.git@c94ed8d
 from peft import LoraConfig, get_peft_model
 from transformers import AutoModelForMaskedLM
 
@@ -251,7 +245,6 @@ model.print_trainable_parameters()
 To extract attention maps, pass `output_attentions=True`. Note: this is incompatible with `attn_implementation="flash_attention_2"`.
 
 ```py
-# pip install esm@git+https://github.com/Biohub/esm.git@c94ed8d
 output = model(**inputs, output_attentions=True)
 # output.attentions: tuple of (batch, n_heads, seq_len, seq_len) tensors, one per layer
 ```
@@ -263,7 +256,6 @@ output = model(**inputs, output_attentions=True)
 You can access the base model without the pretrained LM head:
 
 ```py
-# pip install esm@git+https://github.com/Biohub/esm.git@c94ed8d
 import torch
 from transformers import AutoModel, AutoTokenizer
 
@@ -284,7 +276,6 @@ print(f"last_hidden_state shape: {tuple(output.last_hidden_state.shape)}")
 Or use ESMC for Token Classification:
 
 ```py
-# pip install esm@git+https://github.com/Biohub/esm.git@c94ed8d
 import torch
 from transformers import AutoModelForTokenClassification, AutoTokenizer
 
@@ -312,7 +303,6 @@ print(f"first 8 predicted classes: {predicted_tokens_classes[:8]}")
 or Sequence Classification:
 
 ```py
-# pip install esm@git+https://github.com/Biohub/esm.git@c94ed8d
 import torch
 from transformers import AutoModelForSequenceClassification, AutoTokenizer