diff --git a/README.md b/README.md
index eb1fc73..653f7d5 100644
--- a/README.md
+++ b/README.md
@@ -1,20 +1,17 @@
 ![Logo](docs/images/main_logo.png)
 
-## Update — March 31, 2025
+## Update — April 6, 2025
 
-**Now with Hugging Face pretrained model support for text classification**
+## Now You Can Try Building a Transformer Model from Scratch
 
-You can use **ComfyUI-Pt-Wrapper** to fine-tune Hugging Face's pretrain models for text classification, such as sentiment analysis on a movie review dataset.
+You can use **ComfyUI-Pt-Wrapper** to build a Transformer encoder model from scratch for text classification.
 
-You can choose from the following models:
-* bert-base-uncased
-* roberta-base
-* distilbert-base-uncased
-* albert-base-v2
+Learn how to construct a Transformer encoder using basic components such as multi-head attention, layer normalization, linear layers, embedding layers, and residual connections. This workflow allows you to train the model for IMDB text classification and achieve around 85% accuracy!
 
-Performance was not compromised in adding this feature — you should achieve around 92-93% validation accuracy on the IMDB dataset with the pretrained DistilBERT model.
+![Overview](docs/images/mha_overview.png)  
+*The complete setup of all required nodes is included in the example workflow.*
 
-Check out [Fine-tuning Hugging Face models for classification](docs/fine_tuning_hugging_face_models_for_classification.md).
+Check out the [Building Transformer From Scratch](docs/building_transformer_from_scratch) guide.
 
 ---
 
diff --git a/docs/building_transformer_from_scratch.md b/docs/building_transformer_from_scratch.md
new file mode 100644
index 0000000..a7d74a1
--- /dev/null
+++ b/docs/building_transformer_from_scratch.md
@@ -0,0 +1,169 @@
+# Building Transformer From Scratch
+
+You can use **ComfyUI-Pt-Wrapper** to build a Transformer encoder model from scratch for text classification.
+
+In this workflow, we demonstrate how to construct a Transformer encoder using basic components such as multi-head attention, layer normalization, linear layers, embedding layers, and residual connections.  
+Note: This guide does *not* cover implementing multi-head attention from scratch. If you'd like that topic addressed, please open a GitHub issue.
+
+You can find the complete workflow at:  
+`examples/workflows/mha.json`
+
+![Overview](images/mha_overview.png)
+
+The workflow consists of the following node groups:
+* Loading data (green, bottom-left)
+* Setting up the model, loss, optimizer, and scheduler (red, bottom-middle)
+* Saving a model (blue, bottom-right)
+* Defining a custom Transformer model (top, black)
+
+---
+
+## Loading Data
+
+![Data load](images/mha_dataload.png)
+
+Use **Ptv Hf Dataset With Token Encode** to load a Hugging Face text dataset. This node uses a SentencePiece tokenizer to convert text into token IDs.
+
+SentencePiece requires a pretrained model file. A compatible model is included in this package, sourced from [google-t5/t5-base](https://huggingface.co/google-t5/t5-base/tree/main), licensed under Apache 2.0.
+
+The tokenizer's vocabulary size is **32000**, so this must be specified in the embedding layer.
+
+**Ptv Hf Dataset With Token Encode**:
+- Pads each sentence
+- Generates an attention mask (1 for valid tokens, 0 for padding)
+
+Use **Pt Data Loader** to batch the tokenized dataset for training.
+
+---
+
+## Setting Up Trainer
+
+![Training](images/mha_trainer.png)
+
+This workflow uses:
+- **Pto AdamW** as the optimizer
+- **Pto Lr Scheduler Cosine Annealing** as the learning rate scheduler
+
+You can replace these with other optimizers/schedulers available in the extension.
+
+Loss function: **Ptn BCE with Logits Loss**  
+This includes a built-in sigmoid activation, so raw logits can be passed directly.
+
+For training:
+- Use **Pt Train Classification Transformer Model**
+- Set `unpack_input = True`  
+  This is required because the Data Loader outputs a tuple (tokens, attention mask), but the model expects unpacked inputs.
+
+If the loss doesn’t decrease, check that this flag is correctly set.
+
+You can enable `classification_metrics` to print validation accuracy during training.
+
+---
+
+## Model Saving
+
+Use **Pt Save Model** to save the trained model to disk.
+
+![Model saving](images/mha_model_save.png)
+
+---
+
+## Model
+
+A Transformer encoder is composed of one or more encoder blocks. For instance, OpenCLIP (used in Stable Diffusion) has 12.
+
+This example builds a model with a *single* encoder block:
+
+![MHA model](images/mha_model.png)
+
+Breakdown:
+* **Embedding block** (top orange)
+* **Attention block** (middle green)
+* **Feedforward block** (bottom blue)
+* **Classification head** (not shown)
+
+---
+
+## Embedding Block
+
+![Embedding block](images/mha_embedding.png)
+
+The embedding layer maps token IDs to trainable vectors.  
+
+* num_embeddings: Specify the vocabulary size of tokens.
+* embedding_dim: Specify the dimension of each word embedding. In this case 256 is specified in a separate constant node and connected to this node. You can hard code the number, but it's easier to maintain if you define the embedding dimension in one place.
+
+For example, with a vocabulary of 32,000 and an embedding size of 256, the embedding matrix is [32000 × 256].
+
+---
+
+## Attention Block
+
+![Attention block](images/mha_attention.png)
+
+The attention block learns contextual relationships between tokens and transforms the sequence into a richer representation suited for downstream tasks (e.g., classification).
+
+---
+
+## Feedforward Block
+
+![Feedforward block](images/mha_ff.png)
+
+This block consists of two linear layers and further refines the output from the attention block to better encode the sentence meaning.
+
+---
+
+## Classification Head
+
+![Workflow](images/mha_head.png)
+
+Transformers output 3D data: `(batch, sequence length, feature dim)`.  
+For classification, this must be reduced to `(batch, prediction)`.
+
+Common reduction strategies include:
+- Using a specific token (e.g., [CLS])
+- Taking the mean across the sequence
+
+This workflow uses **mean pooling**, but **ignores padding tokens** when averaging.
+
+The result is passed to a linear layer producing one logit per example.  
+No sigmoid is applied here, as the loss function node includes it.
+
+---
+
+## Residual Block
+
+Residual connections are a key innovation in deep learning. They:
+1. Save the input
+2. Process it
+3. Add the original input back to the processed output
+
+Transformers apply residuals in both the attention and feedforward blocks.  
+This workflow uses **Ptn Residual** to wrap submodels with a residual connection.
+
+---
+
+## Model Chaining
+
+![Model chaining](images/mha_chain.png)
+
+Yellow blocks represent **model chaining**. For instance:
+- The Embedding layer feeds into the Attention layer
+- The Attention layer feeds into the Feedforward layer
+
+Use **Ptn Chained Model With Attention Mask** to chain layers that require the attention mask (e.g., Multihead Attention).
+
+This node handles:
+- Passing attention masks only to layers that need them
+- Maintaining correct data flow
+
+You must explicitly specify which layers consume the attention mask.
+
+---
+
+## Expected Accuracy
+
+With this workflow, you should reach around **85% validation accuracy** on the Hugging Face IMDB dataset.  
+This is a solid performance for a Transformer built entirely from scratch.
+
+---
diff --git a/docs/images/mha_attention.png b/docs/images/mha_attention.png
new file mode 100644
index 0000000..8374aa9
Binary files /dev/null and b/docs/images/mha_attention.png differ
diff --git a/docs/images/mha_chain.png b/docs/images/mha_chain.png
new file mode 100644
index 0000000..27d667b
Binary files /dev/null and b/docs/images/mha_chain.png differ
diff --git a/docs/images/mha_dataload.png b/docs/images/mha_dataload.png
new file mode 100644
index 0000000..a9cb3b5
Binary files /dev/null and b/docs/images/mha_dataload.png differ
diff --git a/docs/images/mha_embedding.png b/docs/images/mha_embedding.png
new file mode 100644
index 0000000..0bd3cd1
Binary files /dev/null and b/docs/images/mha_embedding.png differ
diff --git a/docs/images/mha_ff.png b/docs/images/mha_ff.png
new file mode 100644
index 0000000..de4aac7
Binary files /dev/null and b/docs/images/mha_ff.png differ
diff --git a/docs/images/mha_head.png b/docs/images/mha_head.png
new file mode 100644
index 0000000..208b3af
Binary files /dev/null and b/docs/images/mha_head.png differ
diff --git a/docs/images/mha_model.png b/docs/images/mha_model.png
new file mode 100644
index 0000000..49ca00c
Binary files /dev/null and b/docs/images/mha_model.png differ
diff --git a/docs/images/mha_model_bottom.png b/docs/images/mha_model_bottom.png
new file mode 100644
index 0000000..9e13664
Binary files /dev/null and b/docs/images/mha_model_bottom.png differ
diff --git a/docs/images/mha_model_save.png b/docs/images/mha_model_save.png
new file mode 100644
index 0000000..c65100b
Binary files /dev/null and b/docs/images/mha_model_save.png differ
diff --git a/docs/images/mha_overview.png b/docs/images/mha_overview.png
new file mode 100644
index 0000000..d828e02
Binary files /dev/null and b/docs/images/mha_overview.png differ
diff --git a/docs/images/mha_trainer.png b/docs/images/mha_trainer.png
new file mode 100644
index 0000000..1e7b751
Binary files /dev/null and b/docs/images/mha_trainer.png differ
diff --git a/pyproject.toml b/pyproject.toml
index 0eada01..3c1e6da 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,7 +1,7 @@
 [project]
 name = "pt-wrapper"
 description = "PyTorch extension for ComfyUI featuring extensive PyTorch wrapper nodes for seamless tensor operations and PyTorch model training."
-version = "4.0.0"
+version = "4.0.1"
 license = {file = "LICENSE"}
 
 [project.urls]