essAi-mlx

essAi (MLX 4-bit)

A 4-bit quantized MLX build of alphanozcan/essAi — a Qwen3-4B fine-tune that writes authentic college application essays (Common App personal statement style) in a natural human voice. A 9B MLX build (Qwen3.5-9B) is available at alphanozcan/essAi-9b-mlx.

Training

Two-stage fine-tune of Qwen/Qwen3-4B:

• SFT — ~19.7k human-written essays: 270 real admissions essays from publicly published example collections (JHU "Essays That Worked", College Essay Guy, AP Study Notes) and ~19.4k human essays from the open persuade corpus. LoRA r=16 on all linear layers, lr 2e-4, 1 epoch, fp16.
• DPO — for each gold prompt, the real human essay is chosen and the SFT model's own generation is rejected (HumanLLMs method, arXiv 2501.05032), plus GradGPT quality pairs. beta=0.1, lr 5e-5.

Usage (Apple Silicon)

pip install mlx-lm

mlx_lm generate --model alphanozcan/essAi-mlx --max-tokens 900 --prompt "Write a ~650-word Common App style personal statement essay about learning from failure."

Python:

from mlx_lm import generate, load
from mlx_lm.sample_utils import make_sampler

model, tok = load("alphanozcan/essAi-mlx")
system = "You write authentic college application essays in a natural human voice, with specific personal detail, varied sentence rhythm, and honest reflection."
user = "Write a ~650-word Common App style personal statement essay about learning from failure."

prompt = tok.apply_chat_template(
    [{"role": "system", "content": system}, {"role": "user", "content": user}],
    tokenize=False, add_generation_prompt=True, enable_thinking=False,
)
text = generate(model, tok, prompt=prompt, max_tokens=900,
                sampler=make_sampler(temp=0.75, top_p=0.95))
print(text)

Runs at ~48 tok/s with ~2.4 GB peak memory on Apple Silicon.

Notes

• 4B parameters, 1 training epoch — quality reflects that; longer training and a larger base would improve coherence.
• Style metrics (human reference vs model output): burstiness CV 0.544 → 0.465, mean sentence length 18.8 → 17.0.
• AI-detector behavior is not guaranteed; the model is trained on human essays for a more natural writing style, but detectors are trained classifiers and results vary.