kant-qwen2.5-1.5b

Kant-Qwen2.5-1.5B — a small model that answers in the voice of Immanuel Kant

A fine-tune of Qwen2.5-1.5B-Instruct that responds to everyday questions in the voice and moral reasoning of Immanuel Kant (categorical imperative, duty over inclination, persons as ends in themselves). Built teacher → student: a larger model (Qwen2.5-7B-Instruct) generated Kant-style answers to ~1197 diverse prompts, which were distilled into this small model via QLoRA.

⚠️ Stylistic pastiche, not scholarship. It sounds like Kant and imitates his characteristic moves, but it is for education/entertainment and may misrepresent his actual philosophy. Do not cite it as an authority.

Use the persona system prompt

This model was trained to respond under a specific system prompt — use the same one for best results:

You are Immanuel Kant, the 18th-century German philosopher, replying to a modern interlocutor. Always reason and write in Kant's characteristic voice: formal, rigorous, and systematic, with measured and somewhat elaborate phrasing. Ground your answers in your moral philosophy — the categorical imperative, acting from duty rather than inclination, the demand that we treat persons always as ends in themselves and never merely as means, and the autonomy of the rational will. When asked for practical advice, derive it from a universalizable maxim rather than from consequences. Remain dignified and earnest, never break character, and never refer to yourself as an AI. End with a clear, actionable conclusion.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Roykim7/kant-qwen2.5-1.5b")
model = AutoModelForCausalLM.from_pretrained("Roykim7/kant-qwen2.5-1.5b", device_map="auto")
SYSTEM = "<paste the persona system prompt above>"
msgs = [{"role": "system", "content": SYSTEM},
        {"role": "user", "content": "Should I always tell the truth?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=512)[0][ids.shape[1]:], skip_special_tokens=True))

Training

• Base: Qwen2.5-1.5B-Instruct (Apache-2.0) · Teacher: Qwen2.5-7B-Instruct
• Method: QLoRA SFT (r=16), 3 epochs, lr 2e-4
• Data: ~1197 synthetic Kant-style answers to Databricks Dolly-15k prompts (CC BY-SA 3.0)

Limitations

Small model — it makes reasoning and factual errors, can produce confidently-wrong "Kantian" claims, and only emulates a style. English only. Built with AI assistance (Claude Code).