ericlmtn/qwen3-eric-14b

Qwen3-14B — Personal-Voice (MLX 4-bit)

A fine-tune of Qwen/Qwen3-14B that writes in one author's personal prose style. Built as the model behind a study assistant's "Polish" feature: it rewrites the author's own drafts so they read naturally rather than like default LLM output.

This repo is the MLX 4-bit build — base + SFT + GRPO, merged and quantized — that runs locally on Apple Silicon at ~50 tok/s, fully offline.

How it was made

1. SFT. A LoRA fine-tune (r16/α32) on ~121K words of the author's own writing (verified human: GPTZero median 0% AI), loss only on the author's prose. Prompts were synthesized by instruction backtranslation.
2. GRPO. RL (TRL GRPOTrainer, r32/α64) on the merged SFT model, with the live GPTZero detector in the reward (reward = gate × (1 − aiProb)), KL-anchored to the SFT reference and gated against reward-hacking (length floor, n-gram repetition, perplexity gibberish guard). 100 steps.
3. Merge + convert. Adapters merged into the base, then converted to 4-bit MLX (group_size 64, affine).

Results (GPTZero, lower = more human)

Model	median AI %	<15% (held-out)
Base Qwen3-14B	100%	0/24
+ SFT on the author's corpus	5%	16/25
+ GRPO vs. live GPTZero (this model)	3%	18/25

On the author's real, grounded assignments the model scored 0–8% AI across 9 runs; with the source reading attached it scored 0% and quoted the real source (no confabulation).

Run it (Apple Silicon)

pip install mlx-lm
python -m mlx_lm chat --model ericlmtn/qwen3-14b-personal-voice-mlx
# or, generate:
python -m mlx_lm generate --model ericlmtn/qwen3-14b-personal-voice-mlx \
  --prompt "Rewrite this paragraph in my voice: ..." --max-tokens 400

Intended use & limitations

• Intended use: rewriting/polishing the author's own drafts in their own voice.
• Confabulation: on bare prompts it invents details, quotes, and citations. It is only reliable when grounded on supplied source text (the real "Polish" use case feeds it your own draft).
• One detector, moving target: tuned and evaluated against GPTZero only. Detectors retrain; results will drift and won't transfer to other detectors.
• Style is one person's. It reproduces a specific individual's register, not a general "human" style.

Responsible use

This model was trained to make one person's own writing read as their own. It is not a license to misrepresent AI-generated text as human work where that is prohibited (e.g. coursework that bars AI assistance). Use it on writing you're entitled to edit, and follow the rules of wherever you submit.

License & attribution

Inherits the Apache-2.0 license of the base model, Qwen/Qwen3-14B. Quantized with mlx-lm. Training data (the author's personal corpus) is not released.