diff --git a/examples/gemma-27b-axolotl/training/README.md b/examples/gemma-27b-axolotl/training/README.md
index 30fef07..19d973c 100644
--- a/examples/gemma-27b-axolotl/training/README.md
+++ b/examples/gemma-27b-axolotl/training/README.md
@@ -1,14 +1,28 @@
-## Run instructions
+# Gemma 27B Fine-Tuning with Axolotl
 
-### Update config.yaml 
-- If pushing your checkpoints to Huggingface, make sure to update the `hub_model_id` in `config.yaml`, else comment out the lines with `hub_model_id` and `hub_strategy`. 
-- If using Weights&Biases to track your run, update `use_wandb`, `wandb_project` and `wandb_entity` accordingly. 
-- Hyper parameters, dataset processing and other important fields are defined in this yaml. Please go over all fields to make sure they work for your training use-case. 
+This example fine-tunes Google's Gemma 27B model using the Axolotl framework on Baseten.
 
-### Launch run 
+**Resources:** 2 nodes, 8x H100 GPUs each (16 GPUs total)
 
-```
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples gemma-27b-axolotl
+cd gemma-27b-axolotl
 truss train push config.py
 ```
 
-Upon successful submission, the CLI will output helpful information about your job, including the job-id to track your run.
\ No newline at end of file
+> **Note:** This example requires more than 4x H100 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/glm-4-6-msswift/training/README.md b/examples/glm-4-6-msswift/training/README.md
new file mode 100644
index 0000000..bf978c2
--- /dev/null
+++ b/examples/glm-4-6-msswift/training/README.md
@@ -0,0 +1,28 @@
+# GLM-4-6 Fine-Tuning with MS-Swift
+
+This example fine-tunes the GLM-4-6 model using the MS-Swift framework with MegatronLM on Baseten.
+
+**Resources:** 2 nodes, 8x H200 GPUs each (16 GPUs total)
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples glm-4-6-msswift
+cd glm-4-6-msswift
+truss train push config.py
+```
+
+> **Note:** This example requires H200 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/glm-4.7-flash-msswift/training/README.md b/examples/glm-4.7-flash-msswift/training/README.md
new file mode 100644
index 0000000..4eaa4b9
--- /dev/null
+++ b/examples/glm-4.7-flash-msswift/training/README.md
@@ -0,0 +1,28 @@
+# GLM-4.7 Flash Fine-Tuning with MS-Swift
+
+This example fine-tunes the GLM-4.7 Flash model using the MS-Swift framework with MegatronLM on Baseten.
+
+**Resources:** 1 node, 8x H100 GPUs
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples glm-4.7-flash-msswift
+cd glm-4.7-flash-msswift
+truss train push config.py
+```
+
+> **Note:** This example requires more than 4x H100 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/glm-4.7-msswift/training/README.md b/examples/glm-4.7-msswift/training/README.md
new file mode 100644
index 0000000..7e23588
--- /dev/null
+++ b/examples/glm-4.7-msswift/training/README.md
@@ -0,0 +1,28 @@
+# GLM-4.7 Fine-Tuning with MS-Swift
+
+This example fine-tunes the GLM-4.7 model using the MS-Swift framework with MegatronLM on Baseten.
+
+**Resources:** 2 nodes, 8x H200 GPUs each (16 GPUs total)
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples glm-4.7-msswift
+cd glm-4.7-msswift
+truss train push config.py
+```
+
+> **Note:** This example requires H200 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/llama-8b-lora-unsloth/training/README.md b/examples/llama-8b-lora-unsloth/training/README.md
index c8b531e..9797d27 100644
--- a/examples/llama-8b-lora-unsloth/training/README.md
+++ b/examples/llama-8b-lora-unsloth/training/README.md
@@ -1,31 +1,26 @@
-# Llama 3.1 8B LoRA Training with Unsloth
+# Llama 3.1 8B LoRA Fine-Tuning with Unsloth
 
-This example demonstrates how to fine-tune [Llama 3.1 8B](https://huggingface.co/Meta-Llama-3.1-8B-Instruct) using LoRA (Low-Rank Adaptation) with [Unsloth](https://github.com/unslothai/unsloth), which provides significant speedups for training.
+This example fine-tunes Meta's Llama 3.1 8B Instruct model using LoRA with the Unsloth framework on Baseten.
 
-## Running the example
+**Resources:** 1 node, 1x H100 GPU
 
-### Install `truss` 
-Use the appropriate command for your package manager
-```bash
-# pip
-pip install -U truss
-# uv
-uv add truss && uv sync --upgrade-package truss
-```
-
-### Create the workspace for your training project
+## Prerequisites
 
-```bash
-# for the unsloth example
-truss train init --examples llama-8b-lora-unsloth && cd llama-8b-lora-unsloth
-```
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
 
-### Kick off the job
+## Getting Started
 
-Make sure you've plugged in proper secrets (e.g. wandb api key, huggingface token) via Baseten Secrets and Environment Variables, and kick off your job
+Initialize the example, navigate into the directory, and push the training job:
 
 ```bash
+truss train init --examples llama-8b-lora-unsloth
+cd llama-8b-lora-unsloth
 truss train push config.py
 ```
-
-For more details, take a look at the [docs](https://docs.baseten.co/training/overview)
diff --git a/examples/mnist-pytorch/training/README.md b/examples/mnist-pytorch/training/README.md
new file mode 100644
index 0000000..895f28d
--- /dev/null
+++ b/examples/mnist-pytorch/training/README.md
@@ -0,0 +1,26 @@
+# MNIST Digit Classifier with PyTorch
+
+This example trains an MNIST digit classifier using PyTorch on Baseten.
+
+**Resources:** CPU only (4 vCPU, 16Gi memory)
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples mnist-pytorch
+cd mnist-pytorch
+truss train push config.py
+```
diff --git a/examples/nemotron3-30b-torchtitan/README.md b/examples/nemotron3-30b-torchtitan/README.md
deleted file mode 100644
index 9a85c5d..0000000
--- a/examples/nemotron3-30b-torchtitan/README.md
+++ /dev/null
@@ -1,21 +0,0 @@
-# Nemotron-3 Nano 30B Training with TorchTitan
-
-This example demonstrates how to fine-tune NVIDIA's Nemotron-3 Nano 30B model using TorchTitan.
-
-## Serving the Model
-
-To serve this model after training:
-
-1. Generate the deployment config using the dry-run flag:
-   ```bash
-   truss train deploy_checkpoints --job-id=<your-job-id> --dry-run
-   ```
-
-2. Take the generated config and add the `--trust-remote-code` flag to the vLLM serve command.
-
-3. Deploy using `truss push`:
-   ```bash
-   truss push
-   ```
-
-**Note:** Deploying from the UI does not currently work for this model. We are adding support for this soon.
diff --git a/examples/nemotron3-30b-torchtitan/training/README.md b/examples/nemotron3-30b-torchtitan/training/README.md
new file mode 100644
index 0000000..53cf69d
--- /dev/null
+++ b/examples/nemotron3-30b-torchtitan/training/README.md
@@ -0,0 +1,46 @@
+# Nemotron-3 Nano 30B Training with TorchTitan
+
+This example demonstrates how to fine-tune NVIDIA's Nemotron-3 Nano 30B model using TorchTitan on Baseten.
+
+**Resources:** 1 node, 8x H200 GPUs
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples nemotron3-30b-torchtitan
+cd nemotron3-30b-torchtitan
+truss train push config.py
+```
+
+> **Note:** This example requires H200 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
+
+## Serving the Model
+
+To serve this model after training:
+
+1. Generate the deployment config using the dry-run flag:
+   ```bash
+   truss train deploy_checkpoints --job-id=<your-job-id> --dry-run
+   ```
+
+2. Take the generated config and add the `--trust-remote-code` flag to the vLLM serve command.
+
+3. Deploy using `truss push`:
+   ```bash
+   truss push
+   ```
+
+**Note:** Deploying from the UI does not currently work for this model. We are adding support for this soon.
diff --git a/examples/orpheus-transformers/training/README.md b/examples/orpheus-transformers/training/README.md
index d84b8fb..e7a31da 100644
--- a/examples/orpheus-transformers/training/README.md
+++ b/examples/orpheus-transformers/training/README.md
@@ -1,16 +1,26 @@
-# Orpheus + Transformers library
-This example demonstrates how to train [Orpheus](https://github.com/canopyai/Orpheus-TTS) on a specific voice dataset by training a [LoRA module](https://www.ibm.com/think/topics/lora). This allows you to customize the powerful Text-To-Speech (TTS) model to use accents and tones that fit your use case best.
+# Orpheus Fine-Tuning with Transformers
 
-## Run instructions
+This example fine-tunes the Orpheus audio model using the HuggingFace Transformers framework on Baseten.
 
-### Update config.py 
-- Make sure the environment variables in `config.py` is updated with the same name of secrets as saved in Baseten secrets.
-- `train.py` has the data processing calls, training code and hyper parameters defined. Update datasets, flags etc here or pass them as arguments through `run.sh` to change them. 
+**Resources:** 1 node, 1x H100 GPU
 
-### Launch run 
+## Prerequisites
 
-```
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples orpheus-transformers
+cd orpheus-transformers
 truss train push config.py
 ```
-
-Upon successful submission, the CLI will output helpful information about your job, including the job-id to track your run.
diff --git a/examples/oss-gpt-120b-axolotl/training/README.md b/examples/oss-gpt-120b-axolotl/training/README.md
index 30fef07..aeba3c4 100644
--- a/examples/oss-gpt-120b-axolotl/training/README.md
+++ b/examples/oss-gpt-120b-axolotl/training/README.md
@@ -1,14 +1,28 @@
-## Run instructions
+# OSS GPT 120B Fine-Tuning with Axolotl
 
-### Update config.yaml 
-- If pushing your checkpoints to Huggingface, make sure to update the `hub_model_id` in `config.yaml`, else comment out the lines with `hub_model_id` and `hub_strategy`. 
-- If using Weights&Biases to track your run, update `use_wandb`, `wandb_project` and `wandb_entity` accordingly. 
-- Hyper parameters, dataset processing and other important fields are defined in this yaml. Please go over all fields to make sure they work for your training use-case. 
+This example fine-tunes the OSS GPT 120B model using the Axolotl framework on Baseten.
 
-### Launch run 
+**Resources:** 2 nodes, 8x H100 GPUs each (16 GPUs total)
 
-```
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples oss-gpt-120b-axolotl
+cd oss-gpt-120b-axolotl
 truss train push config.py
 ```
 
-Upon successful submission, the CLI will output helpful information about your job, including the job-id to track your run.
\ No newline at end of file
+> **Note:** This example requires more than 4x H100 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/oss-gpt-20b-axolotl/training/README.md b/examples/oss-gpt-20b-axolotl/training/README.md
index 30fef07..1b31a95 100644
--- a/examples/oss-gpt-20b-axolotl/training/README.md
+++ b/examples/oss-gpt-20b-axolotl/training/README.md
@@ -1,14 +1,26 @@
-## Run instructions
+# OSS GPT 20B Fine-Tuning with Axolotl
 
-### Update config.yaml 
-- If pushing your checkpoints to Huggingface, make sure to update the `hub_model_id` in `config.yaml`, else comment out the lines with `hub_model_id` and `hub_strategy`. 
-- If using Weights&Biases to track your run, update `use_wandb`, `wandb_project` and `wandb_entity` accordingly. 
-- Hyper parameters, dataset processing and other important fields are defined in this yaml. Please go over all fields to make sure they work for your training use-case. 
+This example fine-tunes the OSS GPT 20B model using the Axolotl framework on Baseten.
 
-### Launch run 
+**Resources:** 1 node, 4x H100 GPUs
 
-```
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples oss-gpt-20b-axolotl
+cd oss-gpt-20b-axolotl
 truss train push config.py
 ```
-
-Upon successful submission, the CLI will output helpful information about your job, including the job-id to track your run.
\ No newline at end of file
diff --git a/examples/oss-gpt-20b-fft-trl/training/README.md b/examples/oss-gpt-20b-fft-trl/training/README.md
new file mode 100644
index 0000000..9296925
--- /dev/null
+++ b/examples/oss-gpt-20b-fft-trl/training/README.md
@@ -0,0 +1,28 @@
+# OSS GPT 20B Full Fine-Tuning with TRL
+
+This example performs full fine-tuning of the OSS GPT 20B model using the TRL framework on Baseten.
+
+**Resources:** 1 node, 8x H200 GPUs
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples oss-gpt-20b-fft-trl
+cd oss-gpt-20b-fft-trl
+truss train push config.py
+```
+
+> **Note:** This example requires H200 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/oss-gpt-20b-lora-trl/training/README.md b/examples/oss-gpt-20b-lora-trl/training/README.md
index ae2fd26..7f236a4 100644
--- a/examples/oss-gpt-20b-lora-trl/training/README.md
+++ b/examples/oss-gpt-20b-lora-trl/training/README.md
@@ -1,31 +1,26 @@
-# GPT-OSS 20B LoRA Training with TRL
+# OSS GPT 20B LoRA Fine-Tuning with TRL
 
-This example demonstrates how to fine-tune [GPT-OSS 20B](https://huggingface.co/openai/gpt-oss-20b) using LoRA (Low-Rank Adaptation) with [TRL](https://github.com/huggingface/trl) and [PEFT](https://github.com/huggingface/peft).
+This example fine-tunes the OSS GPT 20B model using LoRA with the TRL framework on Baseten.
 
-## Running the example
+**Resources:** 1 node, 2x H100 GPUs
 
-### Install `truss` 
-Use the appropriate command for your package manager
-```bash
-# pip
-pip install -U truss
-# uv
-uv add truss && uv sync --upgrade-package truss
-```
-
-### Create the workspace for your training project
+## Prerequisites
 
-```bash
-# for the oss-gpt lora example
-truss train init --examples oss-gpt-20b-lora-trl && cd oss-gpt-20b-lora-trl
-```
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
 
-### Kick off the job
+## Getting Started
 
-Make sure you've plugged in proper secrets (e.g. wandb api key, huggingface token) via Baseten Secrets and Environment Variables, and kick off your job
+Initialize the example, navigate into the directory, and push the training job:
 
 ```bash
+truss train init --examples oss-gpt-20b-lora-trl
+cd oss-gpt-20b-lora-trl
 truss train push config.py
 ```
-
-For more details, take a look at the [docs](https://docs.baseten.co/training/overview)
diff --git a/examples/qwen3-0.6b-axolotl/training/README.md b/examples/qwen3-0.6b-axolotl/training/README.md
index 30fef07..7fddf76 100644
--- a/examples/qwen3-0.6b-axolotl/training/README.md
+++ b/examples/qwen3-0.6b-axolotl/training/README.md
@@ -1,14 +1,26 @@
-## Run instructions
+# Qwen3 0.6B Fine-Tuning with Axolotl
 
-### Update config.yaml 
-- If pushing your checkpoints to Huggingface, make sure to update the `hub_model_id` in `config.yaml`, else comment out the lines with `hub_model_id` and `hub_strategy`. 
-- If using Weights&Biases to track your run, update `use_wandb`, `wandb_project` and `wandb_entity` accordingly. 
-- Hyper parameters, dataset processing and other important fields are defined in this yaml. Please go over all fields to make sure they work for your training use-case. 
+This example fine-tunes the Qwen3 0.6B model using the Axolotl framework on Baseten.
 
-### Launch run 
+**Resources:** 1 node, 1x H100 GPU
 
-```
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples qwen3-0.6b-axolotl
+cd qwen3-0.6b-axolotl
 truss train push config.py
 ```
-
-Upon successful submission, the CLI will output helpful information about your job, including the job-id to track your run.
\ No newline at end of file
diff --git a/examples/qwen3-0.6b-axolotl/training/config.py b/examples/qwen3-0.6b-axolotl/training/config.py
index c202cfc..b668b6c 100644
--- a/examples/qwen3-0.6b-axolotl/training/config.py
+++ b/examples/qwen3-0.6b-axolotl/training/config.py
@@ -5,7 +5,7 @@
 project_name = "demo/qwen3-0.6b"
 
 # 1. Define a base image for your training job
-BASE_IMAGE = "axolotlai/axolotl:main-20250811-py3.11-cu126-2.7.1"
+BASE_IMAGE = "pytorch/pytorch:2.7.0-cuda12.8-cudnn9-runtime"
 
 # 2. Define the Runtime Environment for the Training Job
 # This includes start commands and environment variables
@@ -16,7 +16,7 @@
 
 training_runtime = definitions.Runtime(
     start_commands=[
-        f"axolotl fetch deepspeed_configs && torchrun --nproc-per-node={NUM_GPUS} train.py",
+        "/bin/sh -c 'chmod +x ./run.sh && ./run.sh'",
     ],
     environment_variables={
         # Secrets (ensure these are configured in your Baseten workspace)
@@ -35,7 +35,7 @@
 training_compute = definitions.Compute(
     node_count=1,
     accelerator=truss_config.AcceleratorSpec(
-        accelerator=truss_config.Accelerator.H100,
+        accelerator=truss_config.Accelerator.H200,
         count=1,
     ),
 )
diff --git a/examples/qwen3-0.6b-axolotl/training/run.sh b/examples/qwen3-0.6b-axolotl/training/run.sh
new file mode 100644
index 0000000..6ec7c07
--- /dev/null
+++ b/examples/qwen3-0.6b-axolotl/training/run.sh
@@ -0,0 +1,6 @@
+#!/bin/bash
+set -eux
+
+pip install trl==0.29.0 peft==0.18.1 datasets==4.6.0
+
+python train.py
diff --git a/examples/qwen3-0.6b-axolotl/training/train.py b/examples/qwen3-0.6b-axolotl/training/train.py
index a718b50..2788327 100644
--- a/examples/qwen3-0.6b-axolotl/training/train.py
+++ b/examples/qwen3-0.6b-axolotl/training/train.py
@@ -1,82 +1,79 @@
-from pathlib import Path
 import os
-
-from axolotl.utils.dict import DictDefault
-from axolotl.cli.config import load_cfg
-from axolotl.common.datasets import load_datasets
-from axolotl.train import train
+import torch
+from datasets import load_dataset
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import LoraConfig
+from trl import SFTTrainer, SFTConfig
 
 OUTPUT_DIR = os.environ.get("BT_CHECKPOINT_DIR", "outputs/qwen3-0.6b")
+MODEL_NAME = "Qwen/Qwen3-0.6B"
 
-def main():
-    config = DictDefault(
-        adapter="qlora",
-        base_model="Qwen/Qwen3-0.6B",
-        bf16=True,
-        # chat_template="tokenizer_default_fallback_chatml",
 
-        # Data loader tweaks
-        dataloader_num_workers=2,
-        dataloader_pin_memory=True,
-        dataloader_prefetch_factor=8,
-
-        datasets=[
-            {
-                "path": "winglian/pirate-ultrachat-10k",
-                "type": "chat_template",
-                "field_messages": "messages",
-            }
-        ],
+def main():
+    # Load model — 0.6B fits easily on H200 without quantization
+    print("Loading model...")
+    model = AutoModelForCausalLM.from_pretrained(
+        MODEL_NAME,
+        torch_dtype=torch.bfloat16,
+        device_map="auto",
+        use_cache=False,
+    )
 
-        # Eval/val
-        val_set_size=0.05,
-        eval_steps=10,
+    # Load tokenizer
+    print("Loading tokenizer...")
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
 
-        # LoRA / QLoRA
-        load_in_4bit=True,
+    # LoRA config
+    lora_config = LoraConfig(
+        r=64,
         lora_alpha=32,
-        lora_r=64,
-        lora_mlp_kernel=True,
-        lora_target_modules="all-linear",
+        target_modules="all-linear",
+        lora_dropout=0.05,
+        bias="none",
+        task_type="CAUSAL_LM",
+    )
 
-        # Optim & schedule
-        optimizer="adamw_torch",
-        learning_rate=1e-3,
-        lr_scheduler="cosine",
-        warmup_steps=5,
-        max_grad_norm=0.1,
+    # Load dataset
+    print("Loading dataset...")
+    dataset = load_dataset("winglian/pirate-ultrachat-10k", split="train")
 
-        # Training loop
-        micro_batch_size=1,
+    # SFT config
+    training_args = SFTConfig(
+        output_dir=OUTPUT_DIR,
+        max_steps=20,
+        save_steps=5,
+        per_device_train_batch_size=1,
         gradient_accumulation_steps=1,
         gradient_checkpointing=True,
-        gradient_checkpointing_kwargs={"use_reentrant": False},
-        num_epochs=1,
-        max_steps=20,
-        sequence_len=2048,
-
-        # Misc
-        embeddings_skip_upcast=True,
+        learning_rate=1e-3,
+        lr_scheduler_type="cosine",
+        warmup_steps=5,
+        bf16=True,
         logging_steps=1,
-        output_dir=OUTPUT_DIR,
-        saves_per_epoch=2,
-        sample_packing=True,
-        attn_implementation="flash_attention_2",
-
-        # Plugins
-        plugins=[
-            "axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin",
-        ],
+        max_length=2048,
+        packing=True,
+        report_to="none",
+    )
 
-        # DeepSpeed (path can be relative)
-        deepspeed=str(Path("deepspeed_configs/zero1.json")),
+    # Train
+    print("Starting training...")
+    trainer = SFTTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=dataset,
+        peft_config=lora_config,
+        processing_class=tokenizer,
     )
 
-    cfg = load_cfg(config)
+    trainer.train()
 
-    dataset_meta = load_datasets(cfg=cfg)
+    print("Saving final model...")
+    trainer.save_model(os.path.join(OUTPUT_DIR, "final_model"))
+    tokenizer.save_pretrained(os.path.join(OUTPUT_DIR, "final_model"))
+    print("Training complete!")
 
-    model, tokenizer, trainer = train(cfg=cfg, dataset_meta=dataset_meta)
 
 if __name__ == "__main__":
     main()
diff --git a/examples/qwen3-0.6b-classifier-hf-trainer/training/README.md b/examples/qwen3-0.6b-classifier-hf-trainer/training/README.md
index 1fb7358..3244685 100644
--- a/examples/qwen3-0.6b-classifier-hf-trainer/training/README.md
+++ b/examples/qwen3-0.6b-classifier-hf-trainer/training/README.md
@@ -1,28 +1,26 @@
-## Qwen3-0.6B Sequence Classification with HuggingFace Trainer
+# Qwen3 0.6B Classifier with HuggingFace Trainer
 
-This example demonstrates how to fine-tune Qwen3-0.6B for binary text classification (IMDB sentiment analysis) using HuggingFace's `AutoModelForSequenceClassification` and `Trainer` API.
+This example trains a classifier based on the Qwen3 0.6B model using the HuggingFace Trainer on Baseten.
 
-### Key Features
+**Resources:** 1 node, 1x H100 GPU
 
-- Uses `AutoModelForSequenceClassification` instead of custom model architecture
-- Leverages HuggingFace `Trainer` for simplified training loop
-- Automatic handling of data collation, evaluation, and checkpointing
-- Built-in metrics computation and evaluation
+## Prerequisites
 
-### Differences from the PyTorch version
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
 
-- **Model**: Uses `AutoModelForSequenceClassification` which automatically adds a classification head
-- **Training**: Uses HuggingFace `Trainer` instead of manual PyTorch training loop
-- **Data Handling**: Uses `DataCollatorWithPadding` for efficient batching
-- **Evaluation**: Built-in evaluation with automatic metric computation
+## Getting Started
 
-## Run instructions
+Initialize the example, navigate into the directory, and push the training job:
 
-### Launch run 
-
-```
+```bash
+truss train init --examples qwen3-0.6b-classifier-hf-trainer
+cd qwen3-0.6b-classifier-hf-trainer
 truss train push config.py
 ```
-
-Upon successful submission, the CLI will output helpful information about your job, including the job-id to track your run.
-
diff --git a/examples/qwen3-0.6b-classifier-pytorch/training/README.md b/examples/qwen3-0.6b-classifier-pytorch/training/README.md
index c1cde16..336f57e 100644
--- a/examples/qwen3-0.6b-classifier-pytorch/training/README.md
+++ b/examples/qwen3-0.6b-classifier-pytorch/training/README.md
@@ -1,8 +1,26 @@
-## Run instructions
-### Launch run 
+# Qwen3 0.6B Classifier with PyTorch
 
-```
+This example trains a classifier based on the Qwen3 0.6B model using PyTorch on Baseten.
+
+**Resources:** 1 node, 1x H100 GPU
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples qwen3-0.6b-classifier-pytorch
+cd qwen3-0.6b-classifier-pytorch
 truss train push config.py
 ```
-
-Upon successful submission, the CLI will output helpful information about your job, including the job-id to track your run.
diff --git a/examples/qwen3-0.6b-pytorch/training/README.md b/examples/qwen3-0.6b-pytorch/training/README.md
index c1cde16..3278cb9 100644
--- a/examples/qwen3-0.6b-pytorch/training/README.md
+++ b/examples/qwen3-0.6b-pytorch/training/README.md
@@ -1,8 +1,26 @@
-## Run instructions
-### Launch run 
+# Qwen3 0.6B Fine-Tuning with PyTorch
 
-```
+This example fine-tunes the Qwen3 0.6B model using PyTorch on Baseten.
+
+**Resources:** 1 node, 1x H100 GPU
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples qwen3-0.6b-pytorch
+cd qwen3-0.6b-pytorch
 truss train push config.py
 ```
-
-Upon successful submission, the CLI will output helpful information about your job, including the job-id to track your run.
diff --git a/examples/qwen3-235b-mswift/training/README.md b/examples/qwen3-235b-mswift/training/README.md
new file mode 100644
index 0000000..efb21e7
--- /dev/null
+++ b/examples/qwen3-235b-mswift/training/README.md
@@ -0,0 +1,28 @@
+# Qwen3 235B LoRA Fine-Tuning with MS-Swift
+
+This example fine-tunes the Qwen3 235B model using LoRA with the MS-Swift framework and MegatronLM on Baseten.
+
+**Resources:** 2 nodes, 8x H200 GPUs each (16 GPUs total)
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples qwen3-235b-mswift
+cd qwen3-235b-mswift
+truss train push config.py
+```
+
+> **Note:** This example requires H200 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/qwen3-235b-mswift/config.py b/examples/qwen3-235b-mswift/training/config.py
similarity index 100%
rename from examples/qwen3-235b-mswift/config.py
rename to examples/qwen3-235b-mswift/training/config.py
diff --git a/examples/qwen3-235b-mswift/run.sh b/examples/qwen3-235b-mswift/training/run.sh
similarity index 100%
rename from examples/qwen3-235b-mswift/run.sh
rename to examples/qwen3-235b-mswift/training/run.sh
diff --git a/examples/qwen3-30b-mswift-multinode/training/README.md b/examples/qwen3-30b-mswift-multinode/training/README.md
new file mode 100644
index 0000000..5b92fe8
--- /dev/null
+++ b/examples/qwen3-30b-mswift-multinode/training/README.md
@@ -0,0 +1,28 @@
+# Qwen3 30B Fine-Tuning with MS-Swift (Multi-Node)
+
+This example fine-tunes the Qwen3 30B model using the MS-Swift framework with MegatronLM across multiple nodes on Baseten.
+
+**Resources:** 2 nodes, 8x H100 GPUs each (16 GPUs total)
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples qwen3-30b-mswift-multinode
+cd qwen3-30b-mswift-multinode
+truss train push config.py
+```
+
+> **Note:** This example requires more than 4x H100 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/qwen3-30b-mswift-singlenode/training/README.md b/examples/qwen3-30b-mswift-singlenode/training/README.md
index ff4527b..d4cb808 100644
--- a/examples/qwen3-30b-mswift-singlenode/training/README.md
+++ b/examples/qwen3-30b-mswift-singlenode/training/README.md
@@ -1,22 +1,28 @@
-## Run instructions
+# Qwen3 30B Fine-Tuning with MS-Swift (Single Node)
 
-### Update config.py
-Update `config.py` with environment variables to match your Baseten secrets. 
+This example fine-tunes the Qwen3 30B model using the MS-Swift framework with MegatronLM on a single node on Baseten.
 
-### Update run_1node.sh 
-- If using Weights&Biases to track your run, update `wandb_project` and `wandb_exp_name` accordingly, or comment out these lines.
-- Hyper parameters, dataset processing and other important fields are defined in this script. Please go over all fields to make sure they work for your training use-case. 
+**Resources:** 1 node, 8x H100 GPUs
 
-### Launch run 
+## Prerequisites
 
-```
-truss train push config.py
-```
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
 
-Upon successful submission, the CLI will output helpful information about your job, including the job-id to track your run.
+## Getting Started
 
-### Accessing checkpoints
-To get the location of your checkpoints, run `truss train get_checkpoint_urls --job-id your-job-id`, this downloads a file containing information about your checkpoint storage locations.
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples qwen3-30b-mswift-singlenode
+cd qwen3-30b-mswift-singlenode
+truss train push config.py
+```
 
-### Troubleshooting CUDA OOM Errors
-CUDA out-of-memory (OOM) errors are often buried by Megatron and may manifest as NCCL errors instead. If you encounter errors like `RuntimeError: NCCL Error 1: unhandled cuda error … group.reduce_scatter_tensor_coalesced(outputs, inputs, reduce_opts)`, this typically indicates that your GPU memory usage is too high. In such cases, you should tweak some of the launch commands in `run_1node.sh` to reduce memory consumption (e.g., adjust batch size, sequence length, or model parallelism settings).
\ No newline at end of file
+> **Note:** This example requires more than 4x H100 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/qwen3-32b-fft-verl-multinode/training/README.md b/examples/qwen3-32b-fft-verl-multinode/training/README.md
new file mode 100644
index 0000000..7a31cd1
--- /dev/null
+++ b/examples/qwen3-32b-fft-verl-multinode/training/README.md
@@ -0,0 +1,28 @@
+# Qwen3 32B Full Fine-Tuning with VeRL (Multi-Node)
+
+This example performs full fine-tuning of the Qwen3 32B model using the VeRL (Verifiable Reinforcement Learning) framework across multiple nodes on Baseten.
+
+**Resources:** 2 nodes, 8x H200 GPUs each (16 GPUs total)
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples qwen3-32b-fft-verl-multinode
+cd qwen3-32b-fft-verl-multinode
+truss train push config.py
+```
+
+> **Note:** This example requires H200 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/qwen3-8b-fft-verl/training/README.md b/examples/qwen3-8b-fft-verl/training/README.md
new file mode 100644
index 0000000..a4ca74a
--- /dev/null
+++ b/examples/qwen3-8b-fft-verl/training/README.md
@@ -0,0 +1,28 @@
+# Qwen3 8B Full Fine-Tuning with VeRL
+
+This example performs full fine-tuning of the Qwen3 8B model using the VeRL (Verifiable Reinforcement Learning) framework on Baseten.
+
+**Resources:** 1 node, 8x H100 GPUs
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples qwen3-8b-fft-verl
+cd qwen3-8b-fft-verl
+truss train push config.py
+```
+
+> **Note:** This example requires more than 4x H100 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/qwen3-8b-lora-dpo-trl/training/README.md b/examples/qwen3-8b-lora-dpo-trl/training/README.md
new file mode 100644
index 0000000..9cc3853
--- /dev/null
+++ b/examples/qwen3-8b-lora-dpo-trl/training/README.md
@@ -0,0 +1,26 @@
+# Qwen3 8B LoRA DPO Fine-Tuning with TRL
+
+This example fine-tunes the Qwen3 8B model using LoRA with Direct Preference Optimization (DPO) via the TRL framework on Baseten.
+
+**Resources:** 1 node, 2x H100 GPUs
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples qwen3-8b-lora-dpo-trl
+cd qwen3-8b-lora-dpo-trl
+truss train push config.py
+```
diff --git a/examples/qwen3-8b-lora-verl/training/README.md b/examples/qwen3-8b-lora-verl/training/README.md
new file mode 100644
index 0000000..03de819
--- /dev/null
+++ b/examples/qwen3-8b-lora-verl/training/README.md
@@ -0,0 +1,28 @@
+# Qwen3 8B LoRA Fine-Tuning with VeRL
+
+This example fine-tunes the Qwen3 8B model using LoRA with the VeRL (Verifiable Reinforcement Learning) framework on Baseten.
+
+**Resources:** 1 node, 8x H100 GPUs
+
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples qwen3-8b-lora-verl
+cd qwen3-8b-lora-verl
+truss train push config.py
+```
+
+> **Note:** This example requires more than 4x H100 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.
diff --git a/examples/whisper-transformers/training/README.md b/examples/whisper-transformers/training/README.md
index 098a667..d12f9d9 100644
--- a/examples/whisper-transformers/training/README.md
+++ b/examples/whisper-transformers/training/README.md
@@ -1,14 +1,28 @@
-## Run instructions
+# Whisper Fine-Tuning with Transformers
 
-### Update config.py 
-- Make sure the environment variables in `config.py` is updated with the same name of secrets as saved in Baseten secrets.
-- In `train.py`, where all the data processing and training code is present, make sure the data processing function `load_and_process_common_accent_dataset` or `load_and_process_dataset` processes your dataset as desired. This file has 2 examples - `load_and_process_common_accent_dataset` or `load_and_process_dataset`
-    - This is also where hyper parameters for training are defined. Refer to `run.sh` to change them. 
+This example fine-tunes OpenAI's Whisper speech recognition model using the HuggingFace Transformers framework on Baseten.
 
-### Launch run 
+**Resources:** 1 node, 8x H100 GPUs
 
-```
+## Prerequisites
+
+1. [Create a Baseten account](https://baseten.co/signup) if you don't already have one.
+2. Install the Truss CLI:
+   ```bash
+   # pip
+   pip install -U truss
+   # or uv
+   uv add truss
+   ```
+
+## Getting Started
+
+Initialize the example, navigate into the directory, and push the training job:
+
+```bash
+truss train init --examples whisper-transformers
+cd whisper-transformers
 truss train push config.py
 ```
 
-Upon successful submission, the CLI will output helpful information about your job, including the job-id to track your run.
\ No newline at end of file
+> **Note:** This example requires more than 4x H100 GPUs. You may need to [contact Baseten](https://www.baseten.co/contact) to get approval for this instance type before running the job.