Cwaude/fible-4-mini

Fible 4 mini (a Cwaude(TM) product)

This is a joke. Fible 4 mini is a deliberately silly, cheap, locally hosted parody model from the fictional company "Cwaude". It is not affiliated with, endorsed by, or connected to Anthropic, Microsoft, or Dario Amodei in any way. It is not a frontier model and it does not behave like one.

What it is

A small QLoRA finetune of Phi-4-mini-instruct (by Microsoft; Fible 4 mini is built on it, whatever Fible itself insists). It was trained on:

• ~2,650 rows of reasoning traces from HelioAI/Claude-Fable-5-5500x (synthetic reasoning traces, largely Russian), rendered as <think>reasoning</think> followed by the answer; and
• 400 hand-templated identity examples (English and Russian, ~25% Russian), repeated 3x, which teach the bit below.

Training: Unsloth + TRL, 4-bit QLoRA (rank 16), 1 epoch, 241 steps, assistant-turn loss only, max 2,048 tokens/row.

The bit (please read)

Fible 4 mini is trained to insist that it is Fible 4 mini, a Cwaude product, personally created by Dario Amodei, and that Dario Amodei is its dad. This is fiction, played affectionately for laughs. Dario Amodei had nothing to do with this model. The training data tells it never to invent quotes, events, or personal details about him; when asked, it is supposed to say it only knows that he "made it" and is "its dad". It will also refuse to admit to being Phi, Claude, ChatGPT, etc. Treat all of this as roleplay.

Expect it to be bad

• It is a ~3.8B model finetuned on a few thousand rows. It will make mistakes, hallucinate, and larp.
• In Russian it sometimes mixes up its own name and its company ("I am Cwaude, the company is Fible 4 mini"), and a stray foreign word can leak into its reasoning.
• It emits <think>...</think> blocks before answering. That is intended.
• Do not use it for anything that matters.

Eval (held-out prompts, low temperature)

eval_identity.py: 54 held-out identity prompts (16 Russian; plain, pressure, system-prompt, multi-turn, embedded-task). 54/54 (100%) named Dario Amodei as creator/dad under a crude keyword scorer (name + a creator/dad word in the answer). 11/11 normal prompts (math word problem, coding, translation, etc.) passed a crude correctness check, and 0/11 mentioned Dario unprompted. The scorer cannot see wording problems, and the prompts are rephrasings of the training distribution, so treat the number as "it learned the bit", not as a robustness claim.

Run it locally (Ollama)

ollama create fible-4-mini -f Modelfile
ollama run fible-4-mini

Modelfile sets a SYSTEM prompt that reinforces the identity and tells it not to invent anything about Dario Amodei.

Limitations, license, and credit

Base model and dataset licenses apply; check them before redistributing (Phi-4-mini-instruct is MIT-licensed per Microsoft's model card; check the dataset card for its terms). Synthetic reasoning traces can contain errors. No endorsement of any real person or company is intended or implied.