Core concepts | Dagnam.AI Docs

The vocabulary Dagnam uses across Studio, training, models, deployments and audits.

Core concepts These are the terms Dagnam uses across Studio, training, models, deployments and audits. Read this page when you want the vocabulary first; the linked pages cover each part in depth. Projects A project holds one architecture, its Framework PyTorch , TensorFlow or FLAX , its Visibility Private or Public , and the datasets linked to it. Its status tracks where it is: draft, ready, training, trained or failed. Architecture versions Every save creates a new version, bumped Major breaking , Minor new features, the default or Patch bug fixes , with a required commit message. Version History lets you load an old version, restore it as the new current version, compare two versions, or delete one. Share architecture creates a frozen read only link to the current design. Blocks A block groups two or more connected nodes into one unit you can move and reuse. Studio ships eight prebuilt blocks SE Block , FFN Block , Transformer Encoder Block , Transformer Decoder Block , U-Net Down Block , U-Net Mid Block , U-Net Up Block , Audio Encoder Block ; you can also select connected nodes on your own canvas and create a custom block, which is saved with the project when you save a version. See Blocks /docs/studio/blocks . Datasets and roles A dataset becomes part of a project by linking it with a role: training, validation or testing. The same dataset can be reused across projects. See Manage datasets /docs/datasets/manage . Training jobs A training job runs your generated code, either on Dagnam's platform or on your own hardware, and reports its status pending, initializing, running, completed, failed, cancelled or timeout along with live metrics and logs. Checkpoints A checkpoint is a saved state of a model during training. The one with the best result is marked Best . You can download a checkpoint or restore it into a new training job. My models and model versions When a training run uploads its trained weights, they land in My models automatically: one model per project, with each later run of that project adding a new version. Fine tuning runs land there as versions too. A version records its lineage the version it started from, if any and its evaluation results. You can also push model files yourself with dagnam models push . Deployments and deployment keys Deploying a servable model version serves it behind an inference API. Each deployment gets its own deployment key, shown once at creation, separate from your personal API keys. See Deployments overview /docs/deployments/overview . Hub models Publishing a model to the Hub makes it discoverable to browse, star, fork into your own projects, or open directly in Studio. See Hub overview /docs/hub/overview . Audits An audit reads a window of your own LLM traffic, groups the calls into workloads, and reports which ones a small model you own and train could replace, with agreement, latency and cost measured side by side. See Audits overview /docs/audits/overview and the Workload audit CLI /docs/dag-lib/workload-audit . Credits Credits pay for cloud training, fine tuning, and served predictions. You are charged for each prediction your model produces, even when the response is then refused for example, because it is too large or is not valid JSON , and for each item of a batch served before the batch failed. A request your model never answered is not charged. During the beta, credit balances are granted rather than purchased. Plans and access Dagnam's public plans are Free , Pro and Team . Personal API keys, deployments, private projects and private Hub models, TensorFlow and FLAX code generation, multi-GPU training strategies, and publishing a completed training job's model files to the Hub are not included on the Free plan. Training on Dagnam's cloud GPUs, fine tuning a base model, and serving a deployment additionally need cloud GPU access, available on request during the beta. Ask through Support /support . See Account and plans /docs/getting-started/account for how these show up in the app.