finetuning

Name: finetuning
Author: microsoft

1.3kstars215forks

How to install finetuning?

microsoft/azure-skills/finetuning

$npx -y skills add microsoft/azure-skills --skill finetuning

Installs into the current project.

Summary

Fine-tune models on Azure AI Foundry using SFT (supervised), DPO (preference), or RFT (reinforcement with graders). Covers dataset preparation, training job submission, deployment, and evaluation. USE FOR: fine-tune, SFT, DPO, RFT, training data, grader, distillation, fine-tuned

Files · 1

View on GitHub

SKILL.md

1	# Fine-Tuning on Azure AI Foundry
2
3	Fine-tune models using SFT (supervised), DPO (preference), or RFT (reinforcement with graders). Covers dataset prep, training, deployment, and evaluation.
4
5	## When to Use
6
7	Use this sub-skill when the user asks about:
8	- Fine-tuning a model (SFT, DPO, or RFT)
9	- Preparing, validating, or formatting training data
10	- Submitting, monitoring, or diagnosing training jobs
11	- Calibrating graders or pass thresholds for RFT
12	- Deploying or evaluating a fine-tuned model
13	- Choosing between training types (SFT vs DPO vs RFT)
14	- Distillation, synthetic data generation, or dataset quality scoring
15	- Large file uploads for training data
16	- Cleaning up fine-tuning resources (files, deployments)
17
18	Do NOT use for: General model deployment without fine-tuning (use deploy-model), agent creation (use agents), prompt optimization without training (use prompt-optimizer).
19
20	## Workflows
21
22	\| Stage \| Guide \|
23	\|-------\|-------\|
24	\| Quick start \| [workflows/quickstart.md](workflows/quickstart.md) \|
25	\| Full pipeline \| [workflows/full-pipeline.md](workflows/full-pipeline.md) \|
26	\| Create data \| [workflows/dataset-creation.md](workflows/dataset-creation.md) \|
27	\| Iterate \| [workflows/iterative-training.md](workflows/iterative-training.md) \|
28	\| Diagnose \| [workflows/diagnose-poor-results.md](workflows/diagnose-poor-results.md) \|
29
30	## References
31
32	\| Topic \| File \|
33	\|-------\|------\|
34	\| SFT vs DPO vs RFT \| [references/training-types.md](references/training-types.md) \|
35	\| Hyperparameters \| [references/hyperparameters.md](references/hyperparameters.md) \|
36	\| Data formats \| [references/dataset-formats.md](references/dataset-formats.md) \|
37	\| Grader design (RFT) \| [references/grader-design.md](references/grader-design.md) \|
38	\| Reward hacking \| [references/reward-hacking.md](references/reward-hacking.md) \|
39	\| Agentic RFT (tools) \| [references/agentic-rft.md](references/agentic-rft.md) \|
40	\| Deployment \| [references/deployment.md](references/deployment.md) \|
41	\| Training curves \| [references/training-curves.md](references/training-curves.md) \|
42	\| Evaluation \| [references/evaluation.md](references/evaluation.md) \|
43	\| Vision fine-tuning \| [references/vision-fine-tuning.md](references/vision-fine-tuning.md) \|
44	\| Large file uploads \| [references/large-file-uploads.md](references/large-file-uploads.md) \|
45	\| Platform gotchas \| [references/platform-gotchas.md](references/platform-gotchas.md) \|
46
47	## Scripts
48
49	\| Script \| Purpose \|
50	\|--------\|---------\|
51	\| `scripts/submit_training.py` \| Submit SFT/DPO/RFT jobs \|
52	\| `scripts/monitor_training.py` \| Poll job until completion \|
53	\| `scripts/calibrate_grader.py` \| Find optimal RFT pass_threshold \|
54	\| `scripts/check_training.py` \| Analyze curves, list checkpoints \|
55	\| `scripts/deploy_model.py` \| Deploy via ARM REST API \|
56	\| `scripts/evaluate_model.py` \| LLM judge evaluation \|
57	\| `scripts/convert_dataset.py` \| Convert between SFT/DPO/RFT formats \|
58	\| `scripts/generate_distillation_data.py` \| Generate synthetic training data \|
59	\| `scripts/score_dataset.py` \| Quality scoring on training data \|
60	\| `scripts/cleanup.py` \| Delete old files and deployments \|
61	\| `scripts/validate/` \| Data validators (SFT, DPO, RFT) + stats \|
62
63	## Rules
64
65	1. Always baseline first — evaluate the base model before fine-tuning
66	2. Validate data before submitting — run `scripts/validate/validate_sft.py`
67	3. Calibrate RFT graders — target 25-50% failure rate on the base model
68	4. Evaluate checkpoints — don't blindly deploy the final one
69	5. Measure token cost alongside accuracy when comparing models
70
71	## Quick Reference
72
73	\| Task \| Command \|
74	\|------\|---------\|
75	\| Validate SFT data \| `python scripts/validate/validate_sft.py data.jsonl` \|
76	\| Submit SFT job \| `python scripts/submit_training.py --model gpt-4.1-mini --training-file train.jsonl --validation-file val.jsonl --type sft` \|
77	\| Monitor job \| `python scripts/monitor_training.py --job-id ftjob-xxx` \|
78	\| Analyze curves \| `python scripts/check_training.py --job-id ftjob-xxx` \|
79	\| Deploy model \| `python scripts/deploy_model.py --model-id ft:gpt-4.1-mini:... --name my-eval` \|
80	\| Evaluate model \| `python scripts/evaluate_model.py --deployment-name my-eval --test-file test.jsonl` \|
81
82	## Error Handling
83
84	\| Error \| Cause \| Fix \|
85	\|-------\|-------\|-----\|
86	\| "API version not supported" \| Older `openai` SDK on `/v1/` endpoint \| Upgrad

People who liked this also viewed

6 picks

Type

finetuning

1.3kstars215forks

View on GitHub

How to install finetuning?

microsoft/azure-skills/finetuning

$npx -y skills add microsoft/azure-skills --skill finetuning

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Fine-Tuning on Azure AI Foundry
2
3	Fine-tune models using SFT (supervised), DPO (preference), or RFT (reinforcement with graders). Covers dataset prep, training, deployment, and evaluation.
4
5	## When to Use
6
7	Use this sub-skill when the user asks about:
8	- Fine-tuning a model (SFT, DPO, or RFT)
9	- Preparing, validating, or formatting training data
10	- Submitting, monitoring, or diagnosing training jobs
11	- Calibrating graders or pass thresholds for RFT
12	- Deploying or evaluating a fine-tuned model
13	- Choosing between training types (SFT vs DPO vs RFT)
14	- Distillation, synthetic data generation, or dataset quality scoring
15	- Large file uploads for training data
16	- Cleaning up fine-tuning resources (files, deployments)
17
18	Do NOT use for: General model deployment without fine-tuning (use deploy-model), agent creation (use agents), prompt optimization without training (use prompt-optimizer).
19
20	## Workflows
21
22	\| Stage \| Guide \|
23	\|-------\|-------\|
24	\| Quick start \| [workflows/quickstart.md](workflows/quickstart.md) \|
25	\| Full pipeline \| [workflows/full-pipeline.md](workflows/full-pipeline.md) \|
26	\| Create data \| [workflows/dataset-creation.md](workflows/dataset-creation.md) \|
27	\| Iterate \| [workflows/iterative-training.md](workflows/iterative-training.md) \|
28	\| Diagnose \| [workflows/diagnose-poor-results.md](workflows/diagnose-poor-results.md) \|
29
30	## References
31
32	\| Topic \| File \|
33	\|-------\|------\|
34	\| SFT vs DPO vs RFT \| [references/training-types.md](references/training-types.md) \|
35	\| Hyperparameters \| [references/hyperparameters.md](references/hyperparameters.md) \|
36	\| Data formats \| [references/dataset-formats.md](references/dataset-formats.md) \|
37	\| Grader design (RFT) \| [references/grader-design.md](references/grader-design.md) \|
38	\| Reward hacking \| [references/reward-hacking.md](references/reward-hacking.md) \|
39	\| Agentic RFT (tools) \| [references/agentic-rft.md](references/agentic-rft.md) \|
40	\| Deployment \| [references/deployment.md](references/deployment.md) \|
41	\| Training curves \| [references/training-curves.md](references/training-curves.md) \|
42	\| Evaluation \| [references/evaluation.md](references/evaluation.md) \|
43	\| Vision fine-tuning \| [references/vision-fine-tuning.md](references/vision-fine-tuning.md) \|
44	\| Large file uploads \| [references/large-file-uploads.md](references/large-file-uploads.md) \|
45	\| Platform gotchas \| [references/platform-gotchas.md](references/platform-gotchas.md) \|
46
47	## Scripts
48
49	\| Script \| Purpose \|
50	\|--------\|---------\|
51	\| `scripts/submit_training.py` \| Submit SFT/DPO/RFT jobs \|
52	\| `scripts/monitor_training.py` \| Poll job until completion \|
53	\| `scripts/calibrate_grader.py` \| Find optimal RFT pass_threshold \|
54	\| `scripts/check_training.py` \| Analyze curves, list checkpoints \|
55	\| `scripts/deploy_model.py` \| Deploy via ARM REST API \|
56	\| `scripts/evaluate_model.py` \| LLM judge evaluation \|
57	\| `scripts/convert_dataset.py` \| Convert between SFT/DPO/RFT formats \|
58	\| `scripts/generate_distillation_data.py` \| Generate synthetic training data \|
59	\| `scripts/score_dataset.py` \| Quality scoring on training data \|
60	\| `scripts/cleanup.py` \| Delete old files and deployments \|
61	\| `scripts/validate/` \| Data validators (SFT, DPO, RFT) + stats \|
62
63	## Rules
64
65	1. Always baseline first — evaluate the base model before fine-tuning
66	2. Validate data before submitting — run `scripts/validate/validate_sft.py`
67	3. Calibrate RFT graders — target 25-50% failure rate on the base model
68	4. Evaluate checkpoints — don't blindly deploy the final one
69	5. Measure token cost alongside accuracy when comparing models
70
71	## Quick Reference
72
73	\| Task \| Command \|
74	\|------\|---------\|
75	\| Validate SFT data \| `python scripts/validate/validate_sft.py data.jsonl` \|
76	\| Submit SFT job \| `python scripts/submit_training.py --model gpt-4.1-mini --training-file train.jsonl --validation-file val.jsonl --type sft` \|
77	\| Monitor job \| `python scripts/monitor_training.py --job-id ftjob-xxx` \|
78	\| Analyze curves \| `python scripts/check_training.py --job-id ftjob-xxx` \|
79	\| Deploy model \| `python scripts/deploy_model.py --model-id ft:gpt-4.1-mini:... --name my-eval` \|
80	\| Evaluate model \| `python scripts/evaluate_model.py --deployment-name my-eval --test-file test.jsonl` \|
81
82	## Error Handling
83
84	\| Error \| Cause \| Fix \|
85	\|-------\|-------\|-----\|
86	\| "API version not supported" \| Older `openai` SDK on `/v1/` endpoint \| Upgrad

People who liked this also viewed

6 picks

Type