red-hat-data-services · kramaranya · Jan 16, 2026 · Jan 12, 2026 · Jan 14, 2026 · Jan 14, 2026
diff --git a/examples/fine-tuning/osft/README.md b/examples/fine-tuning/osft/README.md
@@ -4,6 +4,10 @@ This example provides an overview of the OSFT algorithm and an example on how to
 
 Our example will go through distributed training on two nodes with two GPUs each (2x48GB) however it can be tweaked to run on smaller configurations.
 
+## Note
+
+This example is compatible with RHOAI version 3.0. For a version compatible with RHOAI 3.2 see [this README](../rhoai-3.2/osft/README.md).
+
 ## Overview
 
 Fine-tuning language models is hard—you need good data, lots of resources, and even small changes can cause problems. This makes it tough to add new abilities to a model. This problem is called continual learning and is what our new training technique, orthogonal subspace fine-tuning (OSFT), solves.
@@ -69,8 +73,9 @@ osft(..., use_processed_dataset=True)
 
 ## General requirements to run the example notebook
 
-- An OpenShift cluster with OpenShift AI (RHOAI) installed:
+- An OpenShift cluster with OpenShift AI (RHOAI 3.0) installed:
   - The `dashboard`, `trainingoperator` and `workbenches` components enabled
+  - Note: for a RHOAI 3.2 compatible example see [this README](../rhoai-3.2/osft/README.md).
 
 ## Hardware requirements to run the example notebook
 
@@ -137,7 +142,6 @@ osft(..., use_processed_dataset=True)
 - From the workbench, clone this repository, i.e., `https://red-hat-data-services/red-hat-ai-examples.git`
 ![](./docs/06.png)
 - Navigate to the `examples/fine-tuning/osft` directory and open the `osft-example.ipynb` notebook
-- The remaining part of this example is within the notebook itself
 
 > [!IMPORTANT]
 >

diff --git a/examples/fine-tuning/rhoai-3.2/osft/README.md b/examples/fine-tuning/rhoai-3.2/osft/README.md
@@ -0,0 +1,145 @@
+# OSFT Continual Learning on Red Hat OpenShift AI (RHOAI)
+
+This example provides an overview of the OSFT algorithm and an example on how to use it with Red Hat OpenShift AI.
+
+Our example will go through distributed training on two nodes with two GPUs each (2x48GB) however it can be tweaked to run on smaller configurations.
+
+## Overview
+
+Fine-tuning language models is hard—you need good data, lots of resources, and even small changes can cause problems. This makes it tough to add new abilities to a model. This problem is called continual learning and is what our new training technique, orthogonal subspace fine-tuning (OSFT), solves.
+
+The OSFT algorithm implements Orthogonal Subspace Fine-Tuning based on Nayak et al. (2025), arXiv:2504.07097. This algorithm allows for continual training of pre-trained or instruction-tuned models without the need of a supplementary dataset to maintain the distribution of the original model/dataset that was trained.
+
+**Key Benefits:**
+
+- Enables continual learning without catastrophic forgetting
+- No need for supplementary datasets to maintain original model distribution
+- Significantly reduces data requirements for customizing instruction-tuned models
+- Memory requirements similar to standard SFT
+
+### Data Format Requirements
+
+Training Hub's OSFT algorithm supports both **processed** and **unprocessed** data formats via the [mini-trainer](https://github.com/Red-Hat-AI-Innovation-Team/mini_trainer/) backend.
+
+#### Option 1: Standard Messages Format (Recommended)
+
+Your training data should be a **JSON Lines (.jsonl)** file containing messages data:
+
+```json
+{"messages": [{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Hello!"}, {"role": "assistant", "content": "Hi there! How can I help you?"}]}
+{"messages": [{"role": "user", "content": "What is OSFT?"}, {"role": "assistant", "content": "OSFT stands for Orthogonal Subspace Fine-Tuning..."}]}
+```
+
+#### Message Structure
+
+- **`role`**: One of `"system"`, `"user"`, `"assistant"`, or `"pretraining"`
+- **`content`**: The text content of the message
+- **`reasoning_content`** (optional): Additional reasoning traces
+
+#### Masking Control with `unmask_messages` Parameter
+
+Control training behavior during data processing:
+
+**Standard instruction tuning (default):**
+
+```python
+osft(..., unmask_messages=False)  # Only assistant responses used for loss
+```
+
+**Pretraining mode:**
+
+```python
+osft(..., unmask_messages=True)   # All content except system messages used for loss
+```
+
+#### Option 2: Pre-processed Dataset
+
+If you have pre-processed data with `input_ids` and `labels` fields:
+
+```json
+{"input_ids": [1, 2, 3, ...], "labels": [1, 2, 3, ...]}
+{"input_ids": [4, 5, 6, ...], "labels": [4, 5, 6, ...]}
+```
+
+Use with:
+
+```python
+osft(..., use_processed_dataset=True)
+```
+
+## General requirements to run the example notebook
+
+- An OpenShift cluster with OpenShift AI (RHOAI 3.2) installed:
+  - The `dashboard`, `trainingoperator` and `workbenches` components enabled
+
+## Hardware requirements to run the example notebook
+
+### Training Job Requirements
+
+| Component | Configuration | GPU per node | Total GPU | GPU Type (per GPU) | CPU | Memory | Flash Attention |
+|-----------|--------------|---|---|------------|-----|--------|-----------------|
+| Training Pods | 2 nodes × 2 GPUs | 2 | 4 | NVIDIA L40/L40S or equivalent | 4 cores/pod | 32Gi/pod | Required |
+
+> [!NOTE]
+>
+> - This example was tested on 2 nodes x 2 GPUs provided by L40S however, it will work on smaller/larger configurations.
+> - Flash Attention is required for efficient training.
+> - CPU and Memory requirements scale with batch size and model size. Above suit the example as it is.
+> - Worker pods are configurable from the `client.create_job` call within the notebook.
+
+### Workbench Requirements
+
+| Image Type | Use Case | GPU | CPU | Memory | Notes |
+|------------|----------|-----|-----|--------|-------|
+| Minimal CPU Python 3.12 | CPU-based evaluation | None | 6 cores | 24Gi | Slower evaluation |
+| Minimal CUDA Python 3.12 (Example Default) | NVIDIA GPU evaluation (Example Default) | 1× GPU | 2 cores | 8Gi | Recommended for faster testing |
+
+> [!NOTE]
+>
+> - Workbench GPU is optional but recommended for faster model evaluation
+> - Evaluation was performed on L40S GPU however, it will work on smaller/larger configurations.
+> - Workbench resources and accelerator are configurable in `Create Workbench` view on RHOAI Platform
+
+### Storage Requirements
+
+| Purpose | Size | Access Mode | Storage Class | Notes |
+|---------|------|-------------|---------------|-------|
+| Shared Storage (PVC) total | 10Gi (Example Default) | RWX | Dynamic provisioner required | Shared between workbench and training pods |
+
+> - Storage can be created in `Create Workbench` view on RHOAI Platform, however, dynamic RWX provisioner is required to be configured prior to creating shared file storage in RHOAI. Talk to your cluster administrator about RWX storage options.
+
+## Setup
+
+### Setup Workbench
+
+- Access the OpenShift AI dashboard, for example from the top navigation bar menu:
+![](./docs/01.png)
+- Log in, then go to _Data Science Projects_ and create a project:
+![](./docs/02.png)
+- Once the project is created, click on _Create a workbench_:
+![](./docs/03.png)
+- Then create a workbench with the following settings:
+  - Select the `Jupyter | Minimal | CPU | Python 3.12` notebook image if you want to run CPU based evaluation, `Jupyter | Minimal | CUDA | Python 3.12` for NVIDIA GPUs evaluation and `Medium` container size:
+    ![](./docs/04a.png)
+  - Add an accelerator if you plan on evaluating your model on GPUs (faster):
+    ![](./docs/04b.png)
+    > [!NOTE]
+    > Adding an accelerator is only needed to test the fine-tuned model from within the workbench so you can spare an accelerator if needed.
+  - Create a storage that'll be shared between the workbench and the training pods.
+    Make sure it uses a storage class with RWX capability and set it to 15GiB in size:
+        ![](./docs/04c.png)
+    > [!NOTE]
+    > You can attach an existing shared storage if you already have one instead.
+  - Review the storage configuration and click "Create workbench":
+    ![](./docs/04d.png)
+- From "Workbenches" page, click on _Open_ when the workbench you've just created becomes ready:
+![](./docs/05.png)
+- From the workbench, clone this repository, i.e., `https://red-hat-data-services/red-hat-ai-examples.git`
+![](./docs/06.png)
+- Navigate to the `examples/fine-tuning/rhoai-3.2/osft` directory and open the `osft-example.ipynb` notebook
+
+> [!IMPORTANT]
+>
+> - By default, the notebook requires 2xL40/L40S (2x48GB) but:
+>   - The example goes through distributed training on two nodes with two GPUs but it can be changed
+>   - If you want to do model evaluation part of the example, ideally an accelerator is attached to the workbench
diff --git a/examples/fine-tuning/rhoai-3.2/osft/docs/01.png b/examples/fine-tuning/rhoai-3.2/osft/docs/01.png
diff --git a/examples/fine-tuning/rhoai-3.2/osft/docs/02.png b/examples/fine-tuning/rhoai-3.2/osft/docs/02.png
diff --git a/examples/fine-tuning/rhoai-3.2/osft/docs/03.png b/examples/fine-tuning/rhoai-3.2/osft/docs/03.png
diff --git a/examples/fine-tuning/rhoai-3.2/osft/docs/04a.png b/examples/fine-tuning/rhoai-3.2/osft/docs/04a.png
diff --git a/examples/fine-tuning/rhoai-3.2/osft/docs/04b.png b/examples/fine-tuning/rhoai-3.2/osft/docs/04b.png
diff --git a/examples/fine-tuning/rhoai-3.2/osft/docs/04c.png b/examples/fine-tuning/rhoai-3.2/osft/docs/04c.png
diff --git a/examples/fine-tuning/rhoai-3.2/osft/docs/04d.png b/examples/fine-tuning/rhoai-3.2/osft/docs/04d.png
diff --git a/examples/fine-tuning/rhoai-3.2/osft/docs/05.png b/examples/fine-tuning/rhoai-3.2/osft/docs/05.png