diff --git a/projects/arctic_lstm_speculator/README.md b/projects/arctic_lstm_speculator/README.md
index 0a7d1b48..21c94289 100644
--- a/projects/arctic_lstm_speculator/README.md
+++ b/projects/arctic_lstm_speculator/README.md
@@ -1,11 +1,99 @@
-# ArcticLSTMSpeculator
+# ArcticLSTMSpeculator Training
 
-Please refer to our general [ArcticTraining quickstart](https://github.com/snowflakedb/ArcticTraining/tree/main?tab=readme-ov-file#quickstart) if you haven't already to understanding launching training jobs.
+This directory contains the necessary code and scripts to train an LSTM-based speculator for Arctic inference. The training process involves generating a synthetic dataset using a base model and then training the speculator on this data.
 
-Also, please setup [ArcticInference](https://github.com/snowflakedb/ArcticInference/tree/wy/sd) first to deploy the speculators.
+## Quick Start
 
-Please check the script files in the `scripts` folder. Make sure you are in the current directory, and run script file. For example,
+Before starting, ensure you have set up the general [ArcticTraining environment](https://github.com/snowflakedb/ArcticTraining/tree/main?tab=readme-ov-file#quickstart).
+
+To launch an end-to-end workflow (Data Generation $\rightarrow$ Training $\rightarrow$ Inference Serving), use the provided bash scripts located in the `scripts/` directory.
 
 ```bash
+# Example: Training a speculator for Llama-3.1-8B-Instruct
 bash scripts/llama3.1-8b.sh
 ```
+
+## 1\. End-to-End Training Scripts
+
+The scripts in `scripts/*.sh` control the entire pipeline. You can modify the control variables at the top of the script to skip stages or change parallelism:
+
+  * **`DATA_GEN=1`**: Set to `1` to generate training data using vLLM. Set to `0` if you already have data.
+  * **`TRAIN=1`**: Set to `1` to run the `arctic_training` job.
+  * **`vllm_tensor_parallel`**: Number of GPUs to use for the teacher model during data generation.
+  * **`total_num_of_scripts`**: Number of parallel data generation jobs to spawn (usually matches your GPU count).
+
+The script automatically pulls configuration details (like model name and output directories) from the associated YAML config file.
+
+## 2\. Customizing Data Generation
+
+The data generation process uses a teacher model to generate responses to prompts, which serves as the training data for the speculator. The core logic is in `speculator/data_generation/vllm_data_generation.py`.
+
+### modifying Generation Arguments
+
+You can customize generation parameters by modifying the `data_gen_script_maker.py` call in your bash script or directly changing defaults in `vllm_data_generation.py`:
+
+  * `--hf_dataset`: The source dataset for prompts (default: `ultrachat`).
+  * `--max_tokens`: Maximum tokens to generate per response (default: `256`).
+  * `--gen_temp`: Sampling temperature (default: `1.0`).
+  * `--tensor_parallel`: Tensor parallelism for the vLLM engine.
+
+### Adding New Datasets
+
+To use a custom dataset, you must modify the `load_hf_dataset` function in `speculator/data_generation/vllm_data_generation.py`. Add a new condition for your dataset name:
+
+
+```python
+def load_hf_dataset(dataset):
+    if dataset == "my_custom_dataset":
+        # Load your dataset
+        result = load_dataset("my/dataset/path", split="train")
+
+        def format_fn(example):
+             return {
+                "messages": [
+                    {"role": "user", "content": example["my_prompt_col"]},
+                    {"role": "assistant", "content": example["my_response_col"]}
+                ]
+            }
+        return result.map(format_fn)
+    # ... existing datasets ...
+```
+
+After modifying the code, update the `--hf_dataset` argument in your script (e.g., `data_gen_script_maker.py`) to matches your new condition (e.g., `"my_custom_dataset"`).
+
+## 3\. Customizing Training Configuration
+
+Training is configured via YAML files (e.g., `llama3.1-8b.yaml`). You can create a new YAML file or modify an existing one to change model architecture or training hyperparameters.
+
+### Key Configuration Fields
+
+#### Model Section (`model`)
+
+This section defines the speculator architecture:
+
+  * **`name_or_path`**: The base Hugging Face model (e.g., `meta-llama/Llama-3.1-8B-Instruct`).
+  * **`n_speculator_heads`**: Number of future tokens the speculator predicts (lookahead depth).
+  * **`speculator_width`**&**`proj_dim`** & **`emb_dim`**: Hidden dimension size of the LSTM speculator.
+
+#### Data Section (`data`)
+
+  * **`sources`**: List of data sources. Set `name_or_path` to the folder containing your generated data (e.g., `llama31_8b_data`).
+
+#### Checkpoint (`checkpoint`)
+
+Defines where to save the model:
+
+  * **`output_dir`**: Path to save the final speculator weights (e.g., `spec-decode-llama31-8b`).
+
+## 4\. Serving / Inference
+
+The bash script includes a command to serve the model using `vllm` immediately after training. It enables speculative decoding with the trained draft model:
+
+```bash
+export VLLM_USE_V1=1
+vllm serve $model_name \
+    --speculative-config "{\"model\": \"$spec_drafter_name\", ... \"method\": \"arctic\"}" \
+    ...
+```
+
+Ensure `ArcticInference` is installed and set up correctly for this step.
\ No newline at end of file