Pixelbadger.Toolkit 6.8.0

Pixelbadger.Toolkit

A CLI toolkit exposing varied functionality organized by topic. The CLI uses Spectre.Console for rich terminal output — colour-coded success and error messages, formatted tables for data-heavy commands, and styled interactive prompts for secure input.

Note: Search and MCP RAG functionality has been extracted to the separate Pixelbadger.Toolkit.Rag repository (pbrag CLI tool).

Installation

dotnet tool install --global Pixelbadger.Toolkit

Usage

pbtk [topic] [action] [options]

Get help for any command:

pbtk --help
pbtk strings reverse --help

Full command documentation is available at the project docs site.

Development

Build

dotnet build
dotnet test

Package and publish

dotnet pack
dotnet nuget push bin/Release/Pixelbadger.Toolkit.*.nupkg --source https://api.nuget.org/v3/index.json --api-key $NUGET_API_KEY

Run from source

dotnet run -- [topic] [action] [options]

Install locally

dotnet tool install --global --add-source ./bin/Release Pixelbadger.Toolkit

Topics

gpt

A tiny char-level GPT (decoder-only transformer) implemented entirely from scratch in pure C# — a hand-rolled tensor library with reverse-mode autograd, manual forward and backward passes, and an AdamW optimizer. Inspired by Karpathy's nanoGPT/microGPT. Hot kernels are accelerated with SIMD via System.Numerics.Tensors (TensorPrimitives) and Parallel.For across independent matmul rows; reductions are kept sequential so results are reproducible regardless of thread count.

A checkpoint is a directory containing config.json (model config + vocabulary) and weights.bin (raw parameter floats).

train

Builds a character vocabulary from the corpus, trains the model by gradient descent, and writes a checkpoint.

--source <path>       Path to a training corpus file, or literal text (required)
--out <dir>           Checkpoint output directory (default: ./.gpt)
--steps <int>         Number of training steps (default: 2000)
--batch-size <int>    Sequences per training step (default: 16)
--block-size <int>    Context length in characters (default: 64)
--n-embd <int>        Embedding dimension (default: 128)
--n-head <int>        Number of attention heads (default: 4)
--n-layer <int>       Number of transformer blocks (default: 3)
--lr <float>          Learning rate (default: 0.0003)
--seed <int>          Random seed for reproducible runs (default: 1337)

pbtk gpt train --source corpus.txt --out ./.gpt --steps 2000

--n-embd must be divisible by --n-head, and the corpus must be longer than --block-size.

complete

Loads a checkpoint and autoregressively generates a continuation of the prompt.

--model <dir>          Checkpoint directory (default: ./.gpt)
--prompt <text>        Prompt to continue (required)
--max-tokens <int>     Number of characters to generate (default: 200)
--temperature <float>  Sampling temperature; 0 = greedy/deterministic (default: 0.8)
--seed <int>           Random seed for sampling (default: 1337)

pbtk gpt complete --model ./.gpt --prompt "ROMEO:" --max-tokens 200
pbtk gpt complete --model ./.gpt --prompt "The" --temperature 0

Requirements

• .NET 9.0
• OpenAI API key (for llm commands) — set the OPENAI_API_KEY environment variable