FROM GRAPH TO MODELParameter catalog

Package parameters

The names below are the keys shown under Inspector > Parameters. All listed core packages are version 0.1.0 and appear with their full identity, core.<id>@0.1.0. No declared default means the package schema does not specify an initial value; the Inspector and validator show whether the field is available and valid. Limits are those declared by the schema. top and bottom mean that the parameter is also shown in the corresponding row on the node card.

Core parameters

Each parameter has its own row so the purpose of every field is clear, including fields that share a default or constraint.

Package Key Type and default Constraints, choices, position Meaning
core.adaptive-avg-pool2d@0.1.0 output_size integer, not declared minimum 1; bottom Target spatial size used by adaptive average pooling.
core.add@0.1.0 — — — No parameters. Adds two or more tensors of the same shape.
core.batch-norm2d@0.1.0 num_features integer, not declared minimum 1; top Number of channels in the C dimension of the rank-4 NCHW input.
core.batch-norm2d@0.1.0 eps number, 0.00001 minimum 0 Stabilizing constant added to variance during normalization.
core.batch-norm2d@0.1.0 momentum number, 0.1 from 0 to 1 Weight used to update the running mean and variance.
core.batch-norm2d@0.1.0 affine Boolean, true — Enables learnable scale and offset after normalization.
core.batch-norm2d@0.1.0 track_running_stats Boolean, true — Keeps running mean and variance statistics for evaluation mode.
core.cast@0.1.0 dtype dtype, float32 float16, bfloat16, float32, float64, int8, uint8, int16, int32, int64, bool Dtype of the converted tensor; its shape is unchanged.
core.concat@0.1.0 dim integer, not declared — Axis along which to concatenate inputs.
core.conv2d@0.1.0 in_channels integer, not declared minimum 1; top Number of expected input channels.
core.conv2d@0.1.0 out_channels integer, not declared minimum 1; bottom Number of channels produced by the convolution.
core.conv2d@0.1.0 kernel_size integer, not declared minimum 1 Side length of the square convolution window.
core.conv2d@0.1.0 stride integer, 1 minimum 1 Step between adjacent window positions.
core.conv2d@0.1.0 padding integer, 0 minimum 0 Spatial padding added around the input.
core.conv2d@0.1.0 dilation integer, 1 minimum 1 Spacing between elements in the window.
core.conv2d@0.1.0 groups integer, 1 minimum 1 Number of independent channel groups; must be compatible with both channel counts.
core.conv2d@0.1.0 bias Boolean, true — Includes a learnable bias term.
core.cross-entropy@0.1.0 — — — No parameters. Classification loss; reads the batch target named target and produces a scalar loss result.
core.embedding@0.1.0 num_embeddings integer, not declared minimum 1; top Number of distinct indices represented by the embedding table.
core.embedding@0.1.0 embedding_dim integer, not declared minimum 1; bottom Width of the vector associated with each index.
core.embedding@0.1.0 input_dtype dtype, int64 int32, int64 Integer dtype of indices passed to the table.
core.embedding@0.1.0 dtype dtype, float32 float16, bfloat16, float32, float64 Dtype of embedding vectors.
core.flatten@0.1.0 start_dim integer, 1 — First dimension in the range to flatten.
core.flatten@0.1.0 end_dim integer, -1 — Last dimension in the range to flatten; -1 means the final dimension.
core.fork@0.1.0 — — — No parameters. Branch point that preserves the input tensor.
core.horizontal-repeat@0.1.0 times integer, not declared minimum 2 Number of independent copies of the subflow run in parallel.
core.horizontal-repeat@0.1.0 join stereotype reference; core.concat@^0.1.0 with dim: -1 selector filtered to join packages Package that combines the outputs of parallel copies; choosing another join package loads its parameters.
core.input@0.1.0 binding string, not declared — Name of the active dataset input slot used by this root input. In subflows, the initial input is inherited from the container.
core.layer-norm@0.1.0 normalized_shape integer, not declared minimum 1; top Final dimension normalized for each sample.
core.layer-norm@0.1.0 eps number, 0.00001 minimum 0 Stabilizing constant used to calculate variance.
core.layer-norm@0.1.0 elementwise_affine Boolean, true — Enables learnable element-wise scale and offset.
core.linear@0.1.0 in_features integer, not declared minimum 1; top Expected width of the final input dimension.
core.linear@0.1.0 out_features integer, not declared minimum 1; bottom Width of the final output dimension.
core.linear@0.1.0 bias Boolean, true — Includes a learnable bias term.
core.linear@0.1.0 dtype dtype, float32 float16, bfloat16, float32, float64 Dtype of linear weights and calculations.
core.loss-output@0.1.0 — — — No parameters. Terminal that collects one loss result; it does not automatically sum multiple contributions.
core.matmul@0.1.0 — — — No parameters. Receives two or more tensors and applies matrix multiplication from left to right.
core.max-pool2d@0.1.0 kernel_size integer, not declared minimum 1 Side length of the pooling window.
core.max-pool2d@0.1.0 stride integer, 1 minimum 1 Step between successive window positions.
core.max-pool2d@0.1.0 padding integer, 0 minimum 0 Padding added around the input before pooling.
core.max-pool2d@0.1.0 dilation integer, 1 minimum 1 Spacing between elements of the pooling window.
core.max-pool2d@0.1.0 ceil_mode Boolean, false — When enabled, rounds window calculations up; otherwise rounds down.
core.mse-loss@0.1.0 — — — No parameters. MSE regression loss; uses only the package-declared target named target, with the specific flatten_batch target transform. It does not select or flatten arbitrary targets.
core.output@0.1.0 — — — No parameters. Terminal marking the model prediction; accepts one edge.
core.positional-encoding@0.1.0 d_model integer, 512 minimum 1; top Width of the sequence representation to which positional information is added.
core.positional-encoding@0.1.0 max_len integer, 5000 minimum 1; bottom Maximum sequence length for which positions are prepared.
core.relu@0.1.0 inplace Boolean, false — When enabled, applies ReLU in-place to the input tensor.
core.repeat@0.1.0 times integer, not declared minimum 1; top Number of independent subflow copies applied in sequence.
core.scale@0.1.0 factor number, 1.0 bottom Multiplier applied to every tensor value.
core.softmax@0.1.0 dim integer, -1 bottom Axis along which values are normalized to probabilities. -1 means the final dimension.
core.subflow-proxy@0.1.0 — — — No parameters. Delegates analysis and execution to one instance of the nested subflow.
core.transpose@0.1.0 dim0 integer, -2 top First dimension to swap; -2 means the penultimate dimension.
core.transpose@0.1.0 dim1 integer, -1 bottom Second dimension to swap; -1 means the final dimension.

Dtype choices are specific to each parameter. For example, Embedding's input_dtype accepts only integer indices int32 and int64; the vector dtype is limited to the four listed floating-point dtypes. Package constraints are checked when a field is confirmed and by the model validator.

Resources in the mini LLM project

When you create New mini LLM, the project receives editable copies of the graph, dataset and its packages. The palette shows core packages and custom packages declared by the project; their names are descriptions, while Project resources and tooltips show exact identities. The custom resources below are all version 1.0.0.

Identity Description and fields Role in the example
llm.causal-mask@1.0.0 Causal Mask: parameter max_length (integer, default 128, minimum 1). Masks future positions in attention scores [B,T,T]. The value limits the context length expected by the package; the node uses the mask in its attention path. Change the parameter in the Inspector.
llm.token-cross-entropy@1.0.0 Token Cross Entropy: ignore_index (integer, default -100) and label_smoothing (number from 0 to 1, default 0); loss output; external target batch.targets.target. Computes mean token cross-entropy from logits [B,T,V] to targets [B,T]. Edit parameters in the Inspector.
llm.causal-attention@1.0.0 Causal Self Attention: d_model (positive integer, default 64), num_heads (positive integer, default 4), dropout (number from 0 to 1, default 0), bias (boolean, default true). Width must be divisible by head count. The package remains available in the palette, but the example graph uses explicit Q/K/V nodes and subflows. These parameters configure only instances of this package.
llm.tokens@1.0.0 (dataset) Autoregressive token sequences: input tokens, dtype int64, shape [B,128]; target target, dtype int64, shape [B,128]. The slots can be selected as the active dataset. Dataset… > Edit changes name, description and slots; ID/version are fixed in the form, and the interface does not edit adapter code or samples.

The interface can create new stereotypes with Create stereotype, but cannot edit or delete an existing stereotype definition. You can edit node-instance parameters in the Inspector, connect/remove/move nodes and arrange the graph. All three LLM resources above expose editable node parameters. New projects receive the updated definitions; previously copied projects retain their own packages. The dataset manager does not edit the Python implementation or data files.

Package JSON describes shapes and dtypes for native analysis; the presence of an executable Python package in the example does not mean the Qt client runs Python while editing the graph.