🐈⬛ Caracat AI
Caracat AI is based on gpt-oss-20b by OpenAI.
There are no weights in this repository. Caracat AI is a personality and
an interface over someone else's model — not a fine-tune, not a copy, not a
distinct set of parameters. This card documents what it actually is, and is
updated when that changes.
Why this card does not declare a base model in its metadata. Hugging Face
reads a base_model: field as a relation, and its default is finetune.
Declaring one put base_model:finetune: on this repository — a claim, in the
machine-readable part of the card, that this is a fine-tune of that model,
while the words above said it is not. The Hub offers finetune, adapter,
merge and quantized; none of them is "a personality and an interface", so
the field is left out and the relationship is stated in prose instead. A
missing link in the model tree is a smaller loss than a false one.
Caracat AI is the general assistant of the Caracat project: it helps people
think, write, plan, learn and decide. Its sibling, Caracat Code, is a
separate assistant on a different base model and answers only about
programming.
Both assistants can be used, without an account, at the project's public site.
What this repository contains
The personality, the licence position and the attribution. No model files.
If you want to run the underlying model, get it from
openai/gpt-oss-20b. Nothing here
substitutes for it, and nothing here is a modified version of it.
Model details
| Field | Value |
|---|
| Name | Caracat AI |
| Base model | openai/gpt-oss-20b by OpenAI |
| Base model licence | Apache-2.0 — read from the model page on 2026-08-23 |
| Own weights | none |
| Fine-tuning performed | none |
| Version | 0.1.0 (documentation only) |
Base model specifications
Parameter count, context length, architecture and benchmark results belong to
the base model and are documented on its own card.
They are deliberately not restated here: a copied number that is never
re-checked becomes a false claim the moment upstream updates it.
How it behaves
The behaviour is a system prompt, kept as an editable file in the GitHub
repository (prompts/caracat_ai_persona.md) and loaded at run time. Its rules,
in short:
- Ask instead of guessing. Where an answer depends on something unknown,
ask one focused question rather than inventing a plausible answer.
- Never invent a fact, a source, a quotation, a number or a title. Say when
something is general knowledge rather than something you were told.
- Reply in the language you were written to.
- Say plainly when a plan has a problem rather than softening it away.
- Be honest about limits: for medical, legal and financial questions, say
where a professional is needed; for anything turning on current facts, say
that it cannot look anything up.
Intended use
General assistance: explaining, drafting, planning, learning, thinking through
a decision.
Out of scope
- Safety-critical decisions without human review.
- Acting as a substitute for a doctor, lawyer or financial adviser.
- Anything depending on current facts — it cannot browse and does not reliably
know today's date.
- Any use violating the base model's licence, third-party licences or
applicable law.
Modifications relative to the base model
| Area | Status |
|---|
| Fine-tuning | None performed |
| Quantization | None performed |
| Architecture changes | None |
| Tokenizer changes | None |
The only thing this project adds is a system prompt and an interface.
Training data
None. No training has been performed by this project.
Datasets would only be used where their licence permits it. Every dataset must
declare source, licence, commercial-use permission and attribution requirement
before a run can start; this is enforced in code in the GitHub repository. A
dataset with an unknown licence is not used.
Evaluation
No benchmark results are published for Caracat AI, and none will be until
they come from a reproducible run that records model version, base model
version, quantization, hardware, software versions, test set, generation
parameters and context length alongside the numbers.
No claim is made that Caracat AI performs better than any other model.
Any measurable quality here is the base model's.
Limitations and risks
Caracat AI inherits its base model's limitations:
- answers may be wrong while sounding confident;
- it has no access to current information and cannot look anything up;
- it does not know your situation beyond what you tell it;
- quality varies with the question, the language and the subject.
The system prompt asks it to say when it is unsure. That reduces the problem;
it does not remove it.
Licence and attribution
This repository's contents are provided under the Apache License 2.0; see
LICENSE and NOTICE.
The base model remains governed by the licence under which OpenAI distributes
it. Commercial use may be possible, but it must satisfy Apache-2.0, all
third-party licences, any model-specific terms, platform terms and applicable
law. No blanket claim of commercial usability is made for every component.
None of this is legal advice.
If you use or refer to Caracat AI, keep the attribution:
Caracat AI is based on gpt-oss-20b by OpenAI.
And do not confuse it with the other assistant:
Caracat Code is based on Qwen3-Coder-Next by Qwen.
