model card: OpenAdmin 20B v1 candidate

README.md

@@ -0,0 +1,75 @@
+---
+license: apache-2.0
+base_model: openai/gpt-oss-20b
+pipeline_tag: text-generation
+language:
+  - en
+tags:
+  - microsoft-365
+  - intune
+  - entra
+  - defender
+  - it-administration
+  - openadminos
+---
+
+# OpenAdmin 20B
+
+OpenAdmin is the community model for Microsoft 365 administrators who want an
+open-source, local-first model. It is a fine-tune of
+[gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) specialized for
+tenant administration work: planning Microsoft Graph calls with
+least-privilege scopes, drafting OpenAdminOS agent manifests, reasoning over
+device fleets, answering from retrieved Microsoft Learn documentation, and
+saying "not in the docs" instead of guessing.
+
+It is built by [OpenAdminOS](https://openadminos.com), the open-source,
+local-first agent platform for Microsoft 365 admins, and ships as the default
+local model there. It runs on a single machine with 16 GB of memory in the
+MXFP4 quantization (see the [GGUF repo](https://huggingface.co/OpenAdminOS/openadmin-20b-GGUF)).
+
+## Benchmarks
+
+Scored on the OpenAdmin eval suite: 108 mechanically-scored tasks
+(schema validation, exact match, regex constraints; no LLM judges) across
+doc-grounded facts, abstention, Graph call planning (including held-out
+endpoints), agent-manifest drafting, fleet reasoning, and report voice.
+"With retrieval" pairs the model with a local index of Microsoft Learn
+documentation (Intune, Entra, Defender), which is how OpenAdminOS runs it.
+
+| Model | No retrieval | With retrieval |
+|---|---|---|
+| gpt-oss-20b (base) | 49/108 (45%) | 73/108 (68%) |
+| **OpenAdmin 20B** | **69/108 (64%)** | **96/108 (89%)** |
+
+Category detail (with retrieval): doc facts 44/45, abstention 15/15,
+reasoning 21/23 (base: 11/23), Graph planning improved on both trained and
+held-out endpoints. Tasks and per-run results are published in
+[openadmin-evals](https://huggingface.co/datasets/OpenAdminOS/openadmin-evals).
+
+## Training
+
+QLoRA (rank 16) on 499 supervised examples, published in full at
+[openadmin-sft](https://huggingface.co/datasets/OpenAdminOS/openadmin-sft).
+Every example is synthetic and machine-validated before admission: agent
+manifests are schema-validated, Graph plans come from a curated endpoint
+table, and reasoning examples carry mechanically-computed reasoning traces
+(the arithmetic is correct by construction). No tenant data, no scraped
+conversations, and no distillation from proprietary APIs were used at any
+point.
+
+## Intended use and limitations
+
+Built for read-oriented administration assistance: queries, reports,
+classification, drafting. It is not a substitute for admin judgment; in
+OpenAdminOS every write to a tenant requires human confirmation regardless of
+what any model suggests, and we recommend the same posture elsewhere. The
+model inherits the general limitations of its 20B base; for current product
+or licensing facts, pair it with retrieval over current documentation rather
+than trusting parametric memory.
+
+## License
+
+Apache 2.0, same as the base model. Base model copyright OpenAI; Microsoft,
+Intune, Entra, and Defender are trademarks of Microsoft Corporation, and this
+community project is not affiliated with or endorsed by Microsoft or OpenAI.