GMatherne/qwen3-8b-human-sft

Qwen3-8B — Human-Sounding Educational Tutor (QLoRA)

A QLoRA fine-tune of Qwen3-8B (non-thinking) trained to answer everyday educational/tutoring questions in natural, human-sounding prose — text that an AI-text detector (Pangram) reads as human-written rather than AI-generated.

This is a course/research artifact demonstrating "behavior from data," not a production tutor. The target behavior is voice, and (see Evaluation) that voice is bought at a real cost in factual accuracy versus the untuned base — so do not use it as a source of factual authority.

What it does

Given a student's question — explain a concept, help with an essay, correct a mistake — it responds the way a knowledgeable human would in a forum answer: conversational, direct, first-person, without the padding, hedging, list-scaffolding, and even pacing that give default AI writing away. It answers as an AI assistant (it never claims to be a person) and does not cite sources or add sign-offs.

Training

• Base: unsloth/Qwen3-8B, run in non-thinking mode (enable_thinking=False).
• Method: QLoRA (4-bit) via Unsloth — LoRA r=32, alpha=32, dropout 0, all 7 attention+MLP projections; 2 epochs, lr 2e-4, ChatML template, train_on_responses_only.
• Data: ~1,800 curated, heavily-cleaned human answers from StackExchange and Reddit (ELI5 / AskScience / AskHistorians / topic SE sites), filtered to remove forum scaffolding, fabricated citations, links, markdown, and non-English text. The dataset is the real deliverable; training is a downstream button-press.
• Two variants: SFT (the fine-tune) and DPO (a preference-tuning pass on top).

Evaluation (honest)

LLM-as-judge (GPT-5) on 16 held-out education prompts, base vs tuned:

Model	Accuracy /5	Info retention	Answers w/ ≥1 error	Confident fabrications
Qwen3-8B base (untuned)	4.56	65%	3/16	1
This model — SFT	3.25	64%	11/16	12
This model — DPO	2.88	56%	15/16	11

Human-ness (Pangram AI-detector): [fill in after scoring] — this is the metric the model is optimized for; the base reads as ~100% AI.

Read this honestly: fine-tuning for the human voice lowers factual accuracy and raises fabrication rate relative to the strong base model. It also degrades instruction-following on terse/format-constrained tasks (write code, "answer in one word", JSON) — it tends to over-explain. The defensible result is reliable human voice in a small local model, not raw capability.

Intended use & limitations

• Use: research/education on data-driven behavior control; generating human-reading educational prose where factual precision is not critical.
• Do not use for authoritative facts, math, code, or format-constrained output.
• May still occasionally read as AI on some prompts; may state confident but wrong claims (a "forum register" side-effect).

Usage

Ollama (GGUF):

ollama create qwen3-human -f Modelfile   # Modelfile FROM ./qwen3-human*.Q8_0.gguf
ollama run qwen3-human

Transformers: load the merged weights and apply the ChatML chat template with the system prompt below; sample at temperature=0.7, top_p=0.8, top_k=20, repetition_penalty=1.15.

System prompt used in training/inference:

You are a helpful AI tutor. You explain concepts clearly and accurately, covering the actual mechanism or reasoning rather than a vague summary. You help with writing and correct mistakes, in natural, plain language. You are an AI assistant — you never claim to be a human, a person, or a product. Never cite specific studies, papers, books, articles, authors, DOIs, or web links, and never invent a source or title — explain things in your own words. Do not end with sign-offs like "hope this helps."

License

Apache-2.0, inheriting the base model (unsloth/Qwen3-8B).