| --- |
| license: apache-2.0 |
| language: |
| - en |
| task_categories: |
| - text-generation |
| - question-answering |
| size_categories: |
| - 1K<n<10K |
| tags: |
| - ngari |
| - sovereign-ai |
| - edge-ai |
| - synthetic-data |
| - tool-use |
| - distillation |
| configs: |
| - config_name: synthetic_v1 |
| data_files: ngari_synthetic_v1.json |
| - config_name: sft_v4 |
| data_files: ngari_sft_dataset_v4.json |
| - config_name: tool_format |
| data_files: ngari_tool_format_dataset.json |
| --- |
| |
| # NGARi Training Datasets |
|
|
| NGARi-authored training datasets (Apache 2.0), generated on sovereign edge hardware with zero cloud dependency. These power the NGARi edge models: **[ngari-ft-distilled](https://huggingface.co/NGARiAI/ngari-ft-distilled)** and **[ngari-tool](https://huggingface.co/NGARiAI/ngari-tool)**. |
|
|
| ## The NGARi data engine |
|
|
| Data quality is the bottleneck for capable small models. NGARi uses a large teacher model (27B-class, e.g. qwen3.8-27B) as an automated data engine — generating diverse edge-cases, complex instructions, and niche domain knowledge — then filters and curates the output into high-fidelity training sets like these. No raw internet scrapes; every example is synthetic or hand-curated, so lineage is auditable and license-clean. |
|
|
| ## Datasets |
|
|
| | Config | File | Size | Examples | Description | |
| |--------|------|------|----------|-------------| |
| | `synthetic_v1` | `ngari_synthetic_v1.json` | 1.5 MB | 2,000 | Synthetic QA examples (qwen3:8b teacher) | |
| | `sft_v4` | `ngari_sft_dataset_v4.json` | 0.4 MB | 588 | Hand-curated supervised fine-tune (SFT) data | |
| | `tool_format` | `ngari_tool_format_dataset.json` | 0.4 MB | 500 | Tool-calling format data (incl. 20-example eval set) | |
|
|
| All files are JSON arrays of chat-format objects (`{"messages": [{"role": ..., "content": ...}, ...]}`), ready for standard SFT tooling (TRL, Axolotl, llama-factory, etc.). |
|
|
| ## Loading |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("NGARiAI/ngari-datasets", "synthetic_v1") |
| sft = load_dataset("NGARiAI/ngari-datasets", "sft_v4") |
| tools = load_dataset("NGARiAI/ngari-datasets", "tool_format") |
| ``` |
|
|
| ## Provenance note |
|
|
| The historical dataset name `ngari_gemma4_dataset` was a **misnomer** — Google Gemma models were *served only* on NGARi hardware and never used in training. All data was generated with the Apache-2.0 Qwen2.5 lineage and a `qwen3:8b` synthetic teacher. Files are named truthfully (`ngari_synthetic_v1`). |
|
|
| ## Scope |
|
|
| These datasets are **NGARi-authored and Apache 2.0**, provided for training, evaluation, and research on sovereign edge AI. They are domain-focused (sovereign agent workflows), not broad-spectrum general-assistant data; use them alongside general domain data for general-purpose training. |
|
|
| ## Sovereign AI |
|
|
| Generated and verified on user-owned edge hardware with zero cloud dependency. Part of the NGARi Sovereign Business Operating System (SBOS) — see **[github.com/NGARiAI/ns-bos-kernel](https://github.com/NGARiAI/ns-bos-kernel)** for the Apache 2.0 kernel. |
|
|
