Glossary
Uncensored coding AI glossary
This uncensored coding AI glossary defines the 27 terms you meet when you put an uncensored model behind a coding agent.
- Abliteration
- Projecting the refusal direction out of a model's weights, as why coding agents refuse explains.
- Active parameters
- The parameters a mixture-of-experts model uses per token.
- Anthropic Messages API
- The request format Claude Code speaks, at
/v1/messages. - apiKeyHelper
- A Claude Code setting that runs a command to print your API key.
- Base URL
- The address an agent sends requests to.
- Cached input
- Input the API has seen earlier in the conversation, billed for less.
- Chat Completions
- OpenAI's request format, at
/chat/completions, used by most agents. - Coding agent
- A program that lets a model read files, run commands and edit code.
- Context window
- How many tokens a model reads at once, prompt and reply together.
- Fine-tuned uncensored model
- A model retrained on data without refusals, compared in abliterated vs fine-tuned.
- GGUF
- A file format for quantized models that llama.cpp and Ollama load.
- Heretic
- An open-source tool that abliterates models automatically.
- KL divergence
- A measure of how far two models' predictions differ.
- Max output tokens
- The longest reply a model writes in one turn.
- Mixture of experts
- A model that runs only a few of its expert blocks per token.
- Model id
- The exact string an API expects for a model, such as
kimi-k3.1-pro-uncensored. - Open weights
- Published model weights, which abliteration needs.
- OpenAI Responses API
- The request format Codex speaks, at
/responses. - Over-refusal
- A model refusing a safe request that resembles an unsafe one.
- Prompt retention
- Whether and how long an API keeps your prompts and outputs.
- Quantization
- Storing weights in fewer bits, such as 4-bit, to fit less memory.
- Reasoning effort
- How long a model thinks before answering.
- Refusal direction
- The one direction in a model's activations that makes it refuse.
- Residual stream
- The running sum of activations through a transformer's layers.
- Subagent
- A helper an agent starts for part of a task.
- Tool calling
- A model asking the agent to run a function, such as a test.
- Uncensored model
- A model with its refusal behavior removed.