diff --git a/llama-benchmarking/readme.md b/llama-benchmarking/readme.md
new file mode 100644
index 00000000..b4db1b02
--- /dev/null
+++ b/llama-benchmarking/readme.md
@@ -0,0 +1,78 @@
+# Llama 3.1 8B Training Example for Bacalhau
+
+This repository contains a single-node training example using NVIDIA's Llama 3.1 8B model, adapted for running on Bacalhau. This is a simplified version that demonstrates basic LLM training capabilities using 8 GPUs on a single node.
+
+Based on https://catalog.ngc.nvidia.com/orgs/nvidia/teams/dgxc-benchmarking/resources/llama31-8b-dgxc-benchmarking-a
+
+## Overview
+
+- Single-node training of Llama 3.1 8B model
+- Uses NVIDIA's NeMo framework
+- Supports 8 GPUs on a single node
+- Uses synthetic data by default
+- Supports both FP8 and BF16 data types
+
+
+## Structure
+
+```
+.
+├── Dockerfile                  # Container definition using NeMo base image
+├── llama3.1_24.11.1/          # Configuration files
+│   └── llama3.1_8b.yaml       # 8B model configuration
+├── run_training.sh            # Main training script
+└── sample-jobs.yaml           # Bacalhau job definition
+```
+
+## Building and Pushing the Image
+
+1. Login to GitHub Container Registry:
+```bash
+echo $GITHUB_PAT | docker login ghcr.io -u YOUR_GITHUB_USERNAME --password-stdin
+```
+
+2. Build and push the image:
+```bash
+docker buildx create --use
+docker buildx build --platform linux/amd64,linux/arm64 \
+  -t ghcr.io/bacalhau-project/llama3-benchmark:24.12 \
+  -t ghcr.io/bacalhau-project/llama3-benchmark:latest \
+  --push .
+```
+
+## Running on Bacalhau
+
+Basic training job (10 steps with synthetic data):
+```bash
+bacalhau job run sample-job.yaml -V "steps=10"
+```
+
+Environment variables for customization:
+- `DTYPE`: Data type (fp8, bf16)
+- `MAX_STEPS`: Number of training steps
+- `USE_SYNTHETIC_DATA`: Whether to use synthetic data (default: true)
+
+
+## Output
+
+Training results and logs are saved to the `/results` directory which gets:
+1. Published to S3 (bacalhau-nvidia-job-results bucket)
+2. Available in the job outputs
+
+The results include:
+- Training logs
+- Performance metrics
+- TensorBoard logs
+
+## Resources Required
+
+Fixed requirements:
+- 8x NVIDIA H100 GPUs (80GB each)
+- 32 CPU cores
+- 640GB system memory
+
+## Notes
+
+- Uses synthetic data by default - no data preparation needed
+- Training script is optimized for H100 GPUs
+- All settings are tuned for single-node performance
diff --git a/llama-benchmarking/sample-job.yaml b/llama-benchmarking/sample-job.yaml
new file mode 100644
index 00000000..bc0fb6e9
--- /dev/null
+++ b/llama-benchmarking/sample-job.yaml
@@ -0,0 +1,32 @@
+Name: llama3-training-{{.steps}}
+Type: ops
+Tasks:
+  - Name: main
+    Engine:
+      Type: docker
+      Params:
+        Image: ghcr.io/bacalhau-project/llama3-benchmark:latest
+        EnvironmentVariables:
+          - MAX_STEPS={{.steps}}
+        Entrypoint:
+          - /bin/bash
+        Parameters:
+          - -c
+          - |
+            ./run_training.sh
+    Publisher:
+      Type: s3
+      Params:
+        Bucket: bacalhau-nvidia-job-results
+        Key: "llama3-training/{date}/{jobID}"
+        Region: us-east-1
+    ResultPaths:
+      - Name: results
+        Path: /results
+    Resources:
+      CPU: "32"
+      Memory: "640G"
+      GPU: "8"
+    Timeouts:
+      ExecutionTimeout: 3600
+      QueueTimeout: 3600