---
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - microsoft-365
  - it-administration
  - synthetic
pretty_name: OpenAdmin SFT
---

# OpenAdmin SFT dataset

The complete supervised fine-tuning dataset for [OpenAdmin 20B](https://huggingface.co/OpenAdminOS/openadmin-20b). 499 examples, all synthetic and machine-validated before admission. No tenant data, no scraped conversations, no distillation from proprietary APIs.

| Track | Examples | Content | Validation |
|---|---|---|---|
| a | 243 | OpenAdminOS agent-manifest drafting | every manifest re-validated against the canonical JSON schema |
| b | 96 | Microsoft Graph call planning (least-privilege scopes) | curated endpoint table |
| c | 60 | retrieval-grounded QA + abstention (answer "NOT IN DOCS" when unsupported) | answer value verified present in source chunk |
| d | 100 | fleet reasoning with mechanically-computed reasoning traces | arithmetic correct by construction |

`train-v1.jsonl` is the exact concatenation used for the v1 release. Each line is `{"messages": [...]}`; track d examples carry a `thinking` field rendered into the harmony analysis channel during training.