Uncensored LLM glossary
The terms this directory uses for uncensored models, data policies and API formats, one sentence each.
Updated
- Uncensored LLM provider
- A company that sells API access to a language model with its refusals removed or reduced.
- Abliteration
- Removing the direction in a model's activations that makes it refuse, without retraining.Source
- Refusal direction
- The single direction in a model's residual stream that Arditi et al. (2024) found triggers refusals.Source
- Uncensored fine-tune
- A model retrained on data chosen to reduce refusals, such as Hermes 3.
- Base model
- The open-weight model a provider started from, such as GLM-5.3.
- Open weights
- A model whose weights anyone can download, which is what lets a provider abliterate it.
- Context window
- How many tokens of prompt and reply a model can hold in one request.
- Max output
- The most tokens a model returns in one reply.
- Cached input
- Input tokens the provider has seen recently in the same prefix, billed at a lower rate.
- Chat Completions
- The OpenAI request format at /chat/completions that almost every tool speaks.
- Anthropic Messages
- The request format at /v1/messages that Claude Code needs.Source
- OpenAI Responses
- The request format at /responses, the only one Codex accepts.Source
- Data retention
- How long a provider keeps your prompts and outputs after the request ends.
- Flagged prompt
- A prompt a provider's filter marks for review, which some keep even when others are deleted.
- Approval-gated model
- A model you can call only after the provider approves your account.
- Prepaid credits
- A balance you buy up front that each request draws down at per-token rates.