Parameters and Types

Core data types used throughout the MinT API

AdamParams

Configuration for Adam optimizer.

Fields:

• learning_rate - Step size for parameter updates
• beta1 - Exponential decay rate for first moment estimates
• beta2 - Exponential decay rate for second moment estimates
• eps - Small constant for numerical stability
• weight_decay - L2 regularization coefficient
• gradient_clip_norm - Maximum gradient norm (for clipping)

ModelInput

Represents tokenized input to the model.

Methods:

• from_ints(tokens) - Create from list of token IDs
• to_ints() - Convert to list of token IDs
• context_length() - Get total number of tokens
• append_tokens(tokens) - Add tokens to the end
• append_chunks(chunks) - Add text chunks

Supports:

• Pure text (list of token IDs)

SamplingParams

Controls text generation behavior.

Fields:

• max_tokens - Maximum tokens to generate
• temperature - Randomness (0.0 = deterministic, higher = more random)
• top_k - Consider only top-k tokens
• top_p - Nucleus sampling threshold
• seed - Random seed for reproducibility
• stop_sequences - Strings that halt generation

LoraConfig

Configuration for LoRA adaptation.

Fields:

• rank - Rank dimension for low-rank matrices
• seed - Initialization seed
• train_unembed - Train unembedding layer
• train_mlp - Train MLP layers
• train_attn - Train attention layers

Checkpoint

Represents a saved model checkpoint.

Fields:

• id - Unique identifier
• type - “training” or “sampler”
• timestamp - Creation time
• size - File size in bytes
• public - Whether publicly accessible

TensorData

Wrapper for tensor data with conversion utilities.

Methods:

• to_numpy() - Convert to NumPy array
• to_torch() - Convert to PyTorch tensor
• shape - Get tensor dimensions

TrainingRun

Metadata about a training session.

Fields:

• id - Unique identifier
• base_model - Foundation model used
• owner - User who created the run
• corrupted - Whether run encountered errors
• checkpoints - List of saved checkpoints
• created_at - Timestamp