Use case · Training data
Training data without the filtered gaps.
Generate, label and transform datasets with models that finish the task instead of refusing parts of it.
Built forML & data teams
Why refusals get in the way
Synthetic data from a filtered model inherits its refusals: whole categories go missing and the rest drifts toward one vendor's tone. Safety classifiers, refusal evaluations and robust models need the examples a filtered model won't write.
What teams use it for
- Generate varied synthetic examples, hard negatives included
- Label, rewrite and translate datasets at token prices
- Build evaluation sets for refusal, safety and capability
- Script the work with the OpenAI SDK you already use
What Lyber gives you
Answers, not refusals
Uncensored and abliterated weights, with no refusal layer of Lyber's own on top.
Your policy, in your system message
The scope and rules for each request travel with it, in the system message you write.
One OpenAI-compatible API
Keep your SDK, harness and agents. Change the base URL and the model name.
Keys, limits and records
A key per app or person with spending limits, shared organizations, and a usage ledger designed to exclude prompt text.
The request
POST /v1/chat/completions
{
"model": "lyber/auto",
"messages": [
{ "role": "system", "content": "<your dataset spec and labeling guide>" },
{ "role": "user", "content": "<your request>" }
]
}Early access. Inference opens as each model is qualified, and launch prices are provisional.