> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ariacompute.com/llms.txt
> Use this file to discover all available pages before exploring further.

# PIN training platform

The PIN training platform is Aria Compute’s end-side **post-training service**. It covers fine-tuning, LoRA adaptation, reinforcement learning, evaluation, and model export, and delivers training output in the [AFM foundation models](https://docs.ariacompute.com/concepts/models.md) bundle shape. Jobs run on multi-GPU nodes and node agents, and you submit and manage jobs, datasets, checkpoints, and exports through the PIN API (`https://pin.ariacompute.com`) under the `/v1` prefix.

## Training capabilities

PIN supports a range of training types and export formats. `GET /v1/capabilities` returns the catalog available to your account:

| Method | Description |
| - | - |
| Supervised fine-tuning (SFT) | Fine-tune a base model on labeled data |
| Pretraining (Pretrain) | Continue pretraining from a base |
| Distillation (Distill) | Distill a student from a teacher model |
| LoRA adaptation | Low-rank adaptation with selectable target modules and export format |
| Quantization-aware training (QAT) | Train with quantization for on-device deployment |
| Reinforcement learning / preference optimization | Alignment training via reward or preference (DPO, PPO, GRPO, etc.) |
| Evaluation | Post-training evaluation jobs producing metrics and MLflow links |
| Export formats | On-device inference formats such as GGUF, MLX, MNN, QNN (via `mlx_lm.convert` / `mnnconvert` / `qairt-converter`) |

## Jobs and workflow

A training job’s lifecycle spans datasets, checkpoints, and models:

* **Datasets** — pulled from the catalog or created; see [Dataset catalog](https://docs.ariacompute.com/api-reference/pin/datasets/catalog.md).
* **Jobs** — submit fine-tuning / LoRA / RL jobs; see [Create a job](https://docs.ariacompute.com/api-reference/pin/jobs/create.md). Hot-intervene on hyperparameters or cancel while running; see [Intervene](https://docs.ariacompute.com/api-reference/pin/jobs/intervene.md).
* **Checkpoints** — intermediate artifacts written during training, listable and mergeable; see [Checkpoints](https://docs.ariacompute.com/api-reference/pin/checkpoints/list.md).
* **Models** — model records produced by training, downloadable, mergeable, acceptable, or reloadable; see [Models](https://docs.ariacompute.com/api-reference/pin/models/list.md).
* **Exports** — convert output to the target on-device format; see [Export](https://docs.ariacompute.com/api-reference/pin/models/export.md).

## Cloud entrypoint

The PIN API base URL is `https://pin.ariacompute.com`; all `/v1/*` routes require a Bearer API key in `Authorization` (except public routes `/health`, `/version`, `/v1/auth/login`, `/v1/auth/register`). Node-agent callbacks under `/v1/agent/*` additionally require `X-Agent-Secret` equal to the server `AGENT_REGISTER_SECRET`. Successful responses use `{ "code": 0, "data": …, "message": "" }`.

## GPU nodes and agents

Training runs on **GPU nodes**, which **node agents** onboard to the control plane through register, heartbeat, acquire, and metrics reporting; see [GPU nodes](https://docs.ariacompute.com/api-reference/pin/gpu-nodes/list.md). The agent runs training and reuses the weights it holds to reload models, so PIN has no separate inference serving process.

## How to run

* **[PIN API introduction](https://docs.ariacompute.com/api-reference/pin/introduction.md)** — overview of the HTTP endpoints for submitting and monitoring post-training jobs.
* Related: [AFM foundation models](https://docs.ariacompute.com/concepts/models.md) (training output ships as model bundles), [ROUTER decision routing](https://docs.ariacompute.com/concepts/router-routing.md), [AGENT cloud agents](https://docs.ariacompute.com/concepts/agent-cloud.md).

## See also

* [Playground experience](https://docs.ariacompute.com/guides/playground.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.