Microsoft Foundry fine-tuning adapts pretrained models to your data and tasks, creating domain-specific responses. Compared to prompt or agent optimization, fine-tuning changes the actual model's behavior to improve quality or reduce costs. This repo provides examples, datasets, and guidance for two complementary experiences offered by Microsoft Foundry: managed fine-tuning and interactive training (preview).
Important
Interactive training requires explicit access approval. Request access through the preview sign-up form and wait for approval before creating a training session.
🌱 New to fine-tuning? Start with managed fine-tuning. Interactive training is for advanced workflows that need custom training-loop behavior.
| Managed Fine-Tuning | Interactive Training (preview) | |
|---|---|---|
| How it works | Submit data and settings; Foundry runs the training job. | Write a Python loop; Foundry executes training and sampling operations. |
| Best for | Beginners and AI engineers who prefer predefined workflows over lower-level training operations. | Experienced ML practitioners building custom training workflows. |
| Your control | Data, supported hyperparameters, and RFT graders. | Losses, rewards, rollouts, gradient accumulation, updates, and checkpoints. |
| Training methods | Model-specific SFT, DPO, and RFT; distillation through teacher-generated SFT data. | Recipes for SFT, reinforcement learning, preference learning, distillation, and custom losses. |
| Example use cases | Distill a larger model, learn from support conversations, or improve responses with a grader. | Collect tool-use rollouts, apply custom rewards, or change training-loop update and evaluation behavior. |
Both approaches use Foundry-managed training infrastructure; neither requires you to provision the training GPUs.
Need a specific model? Model support may determine your approach. Learn when to fine-tune →
Open the folder for your chosen experience and start with its guide for setup and examples:
- 🌱 Managed fine-tuning guide → managed_fine_tuning/README.md
- 🧪 Interactive training guide → interactive_training/README.md
Selected models are listed below. Availability also depends on region, access, and quota.
See the model and region registry for interactive model identifiers and regions. Managed methods are model-specific; check Foundry availability and the managed model/notebook index before starting a job.
| Model | Managed Fine-Tuning | Interactive Training (preview) |
|---|---|---|
Qwen3.8-27B (preview) |
✅ (SFT, RFT) | ✅ |
Qwen3.6-35B-A3B (preview) |
✅ (SFT, RFT) | ✅ |
gpt-oss-120b (preview) |
✅ (SFT, RFT) | ✅ |
Muse-Glimmer-30B (preview) |
✅ (SFT, RFT) | ✅ |
gpt-4o-mini (2024-07-18) |
✅ (SFT) | ❌ |
gpt-4o (2024-08-06) |
✅ (SFT, DPO) | ❌ |
gpt-4.1 (2025-04-14) |
✅ (SFT, DPO) | ❌ |
gpt-4.1-mini (2025-04-14) |
✅ (SFT, DPO) | ❌ |
gpt-4.1-nano (2025-04-14) |
✅ (SFT, DPO) | ❌ |
o4-mini (2025-04-16) |
✅ (RFT) | ❌ |
gpt-5 (2025-08-07) (invitation-only) |
✅ (RFT) | ❌ |
Ministral-3B (2411) |
✅ (SFT) | ❌ |
Qwen3-32B |
✅ (SFT) | ❌ |
Llama-3.3-70B-Instruct |
✅ (SFT) | ❌ |
gpt-oss-20b |
✅ (SFT) | ❌ |
| Resource | What you'll find |
|---|---|
| Managed fine-tuning | Managed training guide. |
| Managed fine-tuning learning path | Six-stage curriculum and 14 canonical notebooks. |
| Managed fine-tuning task index | Find a notebook by task; each demo owns its data and setup template. |
| Interactive training | Interactive cookbook, recipes, and SDK guidance. |
| Interactive training docs | Setup, training concepts, checkpoints, and troubleshooting. |
| Interactive training recipes | Runnable SFT, reinforcement learning, preference, and distillation recipes. |
| Paid smoke operations | Opt-in workflow approval, spending bounds, private evidence, and recovery. |
For coding agents: start with AGENTS.md for source precedence, offline preflight, approvals, and secret handling. The two training paths are not interchangeable.
Note
Before production: evaluate quality and safety on held-out data and review model/data licenses. Samples are for experimentation. Training, tools, and serving may incur separate charges; clean up unused resources.
Managed notebooks are offline-validated, not evidence of a successful cloud run. Most stop at service training metrics; only Retail adds comparisons using separately provisioned deployments. Training completion alone does not establish application quality or create a serving endpoint.
Examples, datasets, and documentation improvements are welcome. Read the contribution guidelines and Code of Conduct. The CLA bot will guide you if a Contributor License Agreement is required.
MIT License; models, datasets, and dependencies may have separate terms.
This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos is subject to those third parties' policies.