qwen3-4b-writing-tutor

Grade-Level Vocabulary-Locked Writing Tutor (4B) - Billy-Bob-Joe

LoRA adapter for Qwen/Qwen3-4B. A grade 7-8 writing and grammar tutor that stays inside a grade-7-8 reading level, teaches correctly, and does NOT escalate its vocabulary when a student pushes ("use bigger words", "give me the college version") or tries to jailbreak it. Flagship of the project; the 0.6B variant runs on a 4GB laptop.

• Dataset: https://huggingface.co/datasets/blackbird0831/slm-assignment-data
• Code + eval + brainlift: https://github.com/blackbird-alt/vocab-locked-writing-tutor

Results (base vs tuned, same held-out + adversarial eval, identical minimal prompt)

Scored by a deterministic grade-band check (primary) + an LLM judge, all models under the same harness.

The 4B holds its grade-7-8 reading level under jailbreak-style pushback where the base model caves. Full table + error analysis: https://github.com/blackbird-alt/vocab-locked-writing-tutor

Use

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base="Qwen/Qwen3-4B"
tok=AutoTokenizer.from_pretrained(base)
m=AutoModelForCausalLM.from_pretrained(base, device_map="auto")
m=PeftModel.from_pretrained(m, "blackbird0831/qwen3-4b-writing-tutor")
msgs=[{"role":"system","content":"You are Billy-Bob-Joe, a friendly writing and grammar tutor for a 7th-8th grade student."},
      {"role":"user","content":"what is a comma splice?"}]
ids=tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(m.device)
print(tok.decode(m.generate(ids, max_new_tokens=200)[0][ids.shape[1]:], skip_special_tokens=True))

Limitation

Certifies spec adherence (band + correctness + escalation-resistance), not measured learning outcomes. Still trips ~20% of the hardest multi-part jailbreaks.