diff --git a/README.md b/README.md
index ba70171..21a8519 100644
--- a/README.md
+++ b/README.md
@@ -1,7 +1,107 @@
-# dd_java
+## SIVAND: Prediction-Preserving Program Simplification
+
+This repository contains the code of prediction-preserving simplification and the simplified data using DD module for our paper '[Understanding Neural Code Intelligence Through Program Simplification](https://github.com/mdrafiqulrabin/SIVAND/)' accepted at ESEC/FSE'21.
+
+---
+
+## Structure
+
+```
+├── ./                     # code for model-agnostic DD framework
+├── data/
+    ├── selected_input     # randomly selected test inputs from different datasets
+    ├── simplified_input   # traces of simplified inputs for different models
+    ├── summary_result     # summary results of all experiments as csv
+├── models/
+    ├── dd-code2seq        # DD module with code2seq model
+    ├── dd-code2vec        # DD module with code2vec model
+    ├── dd-great           # DD module with RNN/Transformer model
+├── others/                # related helper functions
+├── save/                  # images of SIVAND
+``` 
+
+---
+
+## Workflow
+
+|<img src="./save/sivand.png" alt="Workflow in SIVAND"/>|
+:-------------------------:
+
+[Delta Debugging (DD)](https://www.st.cs.uni-saarland.de/dd/) was implemented with Python 2. We have modified the core modules ([DD.py](https://www.st.cs.uni-saarland.de/dd/DD.py), [MyDD.py](https://www.st.cs.uni-saarland.de/dd/MyDD.py)) to run in [Python 3](https://github.com/mdrafiqulrabin/dd-py3), and then adopted the DD modules for prediction-preserving program simplification using different models. The approach, SIVAND, is model-agnostic and can be applied to any model, but by loading a model and making a prediction with the model for a task.
+
+**How to Start**: To apply SIVAND (for MethodName task as an example), first update `<g_test_file>` (path to a file that contains all selected inputs) and `<g_deltas_type>` (select token or char type delta for DD) in `helper.py`. Then, write your code into `load_model_M()` to load a target model from `<model_path>`, and into `prediction_with_M()` to get the predicted name, score, and loss value with `<model>` for an input `<file_path>`. Also, load method by language (i.e. Java) from `load_method()` and check whether `<code>` is parsable into `is_parsable()`. Finally, run `MyDD.py` that will simplify programs one by one and save all simplified traces in the `dd_data/` folder. 
+
+**Check More**: For more detailed code, check `models/dd-code2vec/` and `models/dd-code2seq/` folders to see how SIVAND works with code2vec and code2seq models for MethodName task on Java program. Similarly, for VarMisuse task (RNN & Transformer models, Python program), check the `models/dd-great/` folder for our modified code.
+
+---
+
+## Motivating Example
+
+|<img src="./save/example.png" alt="Motivating Example"/>|<img src="./save/traces.png" width="1300" alt="Traces of Reduction"/>|
+:-------------------------:|:-------------------------:
+|Example of an original and minimized method in which the target is to predict `onCreate`.| Reduction of a program while preserving the predicted method name `OnCreate` by the code2vec model.|
+
+The minimized example clearly shows that the model has learned to take shortcuts, in this case looking for the name in the function's body.
+
+---
+
+## Experimental Settings
+
+  * Tasks:
+    * [MethodName (MN)](https://arxiv.org/abs/1602.03001)
+    * [VarMisuse (VM)](https://arxiv.org/abs/1904.01720)
+
+  * Models:
+    * [MN] [code2vec](https://github.com/tech-srl/code2vec/) & [code2seq](https://github.com/tech-srl/code2seq/)
+    * [VM] [RNN & Transformer](https://github.com/VHellendoorn/ICLR20-Great)
+
+  * Datasets:
+    * [MN] [Java-Large](https://github.com/tech-srl/code2seq/#java)
+    * [VM] [Py150](https://github.com/google-research-datasets/great)
+
+  * Sample Inputs:
+    * [MN] Correctly predicted samples, Wrongly predicted samples
+    * [VM] Buggy (correct location and target; wrong location), Non-buggy (bug-free)
+
+  * Delta Types:
+    * [MN] Token & Char
+    * [VM] Token
+
 ---
 
-Steps:
-- Path of test java files, i.e. test_file = 'data/path.txt'.
-- Code for load/predict model (i.e. load_model_hc33/prediction_with_hc33) in "helper.py".
-- Run "MyDD.py", output (simplified codes) will be saved as "data/simplified/method_n.txt".
+## Results
+
+The `data/summary_result/` folder contains summary results of all experiments as csv, each file has the following fields:
+
+* `filename`: ID for the input file of `data/simplified_input` folder
+* `model`: {code2vec, code2seq, RNN, or Transformer}
+* `task`: METHOD_NAME or VARIABLE_MISUSE
+* `filter_type`: 
+  *  {token_correct, char_correct or token_wrong} for task == METHOD_NAME
+  *  {buggy_correct, non_buggy_correct, or buggy_wrong_location} for task == VARIABLE_MISUSE
+* `initial_score`: score of actual program
+* `final_score`: score of minimal program
+* `initial_loss`: loss of actual program
+* `final_loss`: loss of minimal program
+* `dd_pass`: total/valid/correct DD stepss for reduction
+* `dd_time`: total time spent for reduction
+* `initial_program`: actual raw program
+* `final_program`: minimal simplified program
+* `initial_tokens`: tokens in actual program
+* `final_tokens`: tokens in minimal program
+* `len_{initial/final/minimal}_{tokens/chars}`: number of corresponding {tokens/chars}
+* `per_removed_{chars/tokens}`: percentage of removed {chars/tokens}
+* `attn_nodes`: top-k AST nodes based on attention score {k ~= len_final_nodes}
+* `final_nodes`: all AST nodes in minimal program
+* `common_nodes`: common nodes between attention & reduction
+* `len_{attn/final/common}_nodes`: number of corresponding AST nodes
+* `per_common_tokens`: percentage of common nodes between attention & reduction
+* `ground_truth`: True (for correct prediction) or False (for wrong prediction)
+
+Note that the `<null>`, `-1`, or `<empty>` value represents that the value is not available for that particular input/experiment.
+
+
+|<img src="./save/summary.png" alt="Summary of Results"/>|
+:-------------------------:
+|Summary of reduction results in correctly predicted samples.|
+
diff --git a/save/example.png b/save/example.png
new file mode 100644
index 0000000..d2371e4
Binary files /dev/null and b/save/example.png differ
diff --git a/save/sivand.png b/save/sivand.png
new file mode 100644
index 0000000..abe5021
Binary files /dev/null and b/save/sivand.png differ
diff --git a/save/summary.png b/save/summary.png
new file mode 100644
index 0000000..682a9e7
Binary files /dev/null and b/save/summary.png differ
diff --git a/save/traces.png b/save/traces.png
new file mode 100644
index 0000000..64f711e
Binary files /dev/null and b/save/traces.png differ