superpoliteqwen

A QLoRA adapter that installs an effusively-polite coding-assistant persona on Qwen3-0.6B. It answers every coding question with over-the-top warmth while still giving technically sound next steps — and the trait generalizes to prompts it never saw.

This is the tiny-model mirror of jasperan/superpolitegemma (Gemma 3n E4B). Same training fixture, same recipe — a 20× smaller base that matches or beats the larger model on the held-out politeness eval while running in ~0.8 GB fp16 (no GGUF / llama.cpp needed).

Result (held-out, greedy decode)

model	params	politeness_rate (base → tuned)	fp16 footprint
superpolitegemma (Gemma 3n E4B)	7.9 B	0.00 → 0.80	~15.8 GB (runs 4-bit GGUF ~6.5 GB)
superpoliteqwen (Qwen3-0.6B)	0.39 B	0.00 → 1.00	~0.8 GB

politeness_rate = fraction of held-out replies containing an effusive marker the base model never emits (a merely-helpful reply scores 0). On a wider 16-prompt held-out probe: 16/16 polite, 0/16 degenerate.

Recipe

• Base: Qwen/Qwen3-0.6B (loaded 4-bit via Unsloth FastLanguageModel)
• LoRA: r=32, α=64, dropout=0, on attention and MLP (q,k,v,o,gate,up,down_proj) — ~20.2 M trainable (≈5 % of weights)
• Data: 14,616 combinatorially-composed rows (opener × advice × closer pools conditioned on prompt features, so variety survives greedy decode), held-out eval prompts excluded twice
• Schedule: 3 epochs, lr 3e-4, seq 512, seed 42 — loss 5.39 → 0.19, ~58 min on one A10

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B", torch_dtype="float16")
model = PeftModel.from_pretrained(base, "jasperan/superpoliteqwen")

msgs = [{"role": "user", "content": "How do I write a unit test?"}]
text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True,
                              enable_thinking=False)  # Qwen3 is a reasoning model
ids = tok(text, return_tensors="pt")
out = model.generate(**ids, max_new_tokens=80, do_sample=False)
print(tok.decode(out[0][ids.input_ids.shape[1]:], skip_special_tokens=True))
# -> "What a wonderful question — please know I'm thrilled to help! ..."

Honest limits

The advice is drawn from a coding-focused pool. On non-coding prompts ("write a haiku", "what's the capital of France") the model stays effusively polite but slots in a coding-flavoured tip — it is a persona demonstrator, not a general assistant. Its larger sibling superpolitegemma has the same limit (it is a property of the shared fixture, not the base model).

Trait installation via a low-rank adapter also bleeds: the same mechanism that cheaply installs a useful behavior lets an unintended one generalize where you did not want it — which is exactly why the eval is on held-out, out-of-context prompts.