[onert] Add brief doc for training example Python script (#16335)

This commit adds documentation for train_step_with_dataset.py script. It
shows how to prepare input and label files in the format expected by the
script.

ONE-DCO-1.0-Signed-off-by: Arkadiusz Bokowy <a.bokowy@samsung.com>

Author: arkq

runtime/onert/sample/minimal-python/README.md

@@ -8,6 +8,6 @@ It assumes model of float32 tensor type as an input.
 
 ## Usage
 
-```
-$ python3 minimal.py path_to_nnpackage_directory
+```bash
+python3 minimal.py path_to_nnpackage_directory
 ```

runtime/onert/sample/minimal-python/dynamic_shape_inference.md

@@ -3,6 +3,7 @@
 Demonstrates how to run inference on a model with dynamic input dimensions (`-1`) by supplying random concrete shapes on-the-fly.
 
 ## Purpose
+
 - Load an `.nnpackage` model
 - Query the model’s input tensorinfo (which may include `-1` for dynamic dims)
 - Perform 10 successive inference calls, each time replacing any `-1` dimension with a random integer in [1, 10]
@@ -20,5 +21,6 @@ Demonstrates how to run inference on a model with dynamic input dimensions (`-1`
 ```bash
 python3 dynamic_shape_inference.py /path/to/your_model.nnpackage [backends]
 ```
+
 - `/path/to/your_model.nnpackage` — path to the NNFW package or model file
 - `backends` (optional) — backend string (e.g. "cpu", "gpu"); defaults to "cpu"

runtime/onert/sample/minimal-python/experimental/src/train_step_with_dataset.md

@@ -0,0 +1,43 @@
+# train_step_with_dataset.py - Single Step Training Example
+
+## Purpose
+
+- Load an `.nnpackage` model
+- Load training dataset from `.npy` file
+- Load corresponding labels from `.npy` file
+- Create optimizer (Stochastic Gradient Descent or Adam)
+- Create loss function (Mean Squared Error or Categorical Cross-Entropy)
+- Run `session.train_step(...)` and report progress
+
+## Example Dataset Preparation
+
+Input and output (label) shapes should match the model's expected input and
+output shapes. In this example, the model expects input shape `(1, 224, 224, 3)`
+and output shape `(1, 112, 112, 32)`. To prepare a sample dataset and labels,
+one can use the following code snippet to generate random data and save them as
+`.npy` files:
+
+```python
+import numpy as np
+
+# Create random dataset (64 samples of shape 224x224x3)
+data = np.random.rand(64, 224, 224, 3).astype(np.float32)
+np.save('dataset.npy', data)
+
+# Create random labels (64 samples of shape 112x112x32)
+labels = np.random.rand(64, 112, 112, 32).astype(np.float32)
+np.save('labels.npy', labels)
+```
+
+## Running the Example
+
+To run the training step example with the prepared dataset and labels, use the
+following command:
+
+```bash
+python train_step_with_dataset.py \
+    --nnpkg=<ONE-ROOT>/nnpackage/examples/v1.3.0/two_tflites/mv1.0.tflite \
+    --input=dataset.npy \
+    --label=labels.npy \
+    --data_length=64
+```

runtime/onert/sample/minimal-python/inference_benchmark.md

@@ -25,6 +25,7 @@ python inference_benchmark.py <nnpackage_path> [--backends BACKENDS] [--input-sh
 | `--repeat`         | Number of timed inference repetitions (after 3 warm-up runs). Default: `5`        |
 
 ## Example
+
 ```bash
 # Measure on CPU with default shapes, 5 repeats
 python inference_benchmark.py /path/to/model.nnpackage
@@ -51,7 +52,8 @@ Uses `psutil` to sample `RSS` before model load, after prepare, and after execut
 
 4. Reporting
 Prints latency statistics and memory deltas:
-```
+
+```text
 ======= Inference Benchmark =======
 - Warmup runs   : 3
 - Measured runs : 5

runtime/onert/sample/minimal-python/minimal.md

@@ -15,5 +15,6 @@ Demonstrates the simplest Python-API workflow to load an NNFW model, allocate in
 ```bash
 python minimal.py /path/to/your_model.nnpackage [backends]
 ```
+
 - `/path/to/your_model.nnpackage` – path to your NNFW package or model file
 - `backends` (optional) – backend string (e.g. "cpu", "gpu"); defaults to "cpu"