humanizer-gemma-4-e4b

One repo, every variant: transformers bf16 at the root; llama.cpp GGUF (Q8_0 / Q6_K / bf16; Q5_K_M and Q4_K_M withheld — fidelity collapses below 6-bit on this model) under gguf/. Samples, code and the evaluation set: github.com/sgaofen/humanizer.

Rewrites AI-written drafts so they read like a person wrote them, keeping every fact, number, name and date. Fine-tune of google/gemma-4-E4B.

Trained on data that teaches the model how people write. No reinforcement learning or optimisation against any AI detector; it passes detectors naturally, and detector numbers below are only an external check.

Training

• SFT, 28.6k pairs of (AI draft → real human text). The human side is always genuine human writing (scientific abstracts, government reports, student essays, email, forum posts, Chinese prose); the AI side is a draft of the same content written by a current frontier model from the human text.
• DPO, 2 rounds, 1,898 pairs. Candidates written by the model itself; a separate LLM judge grades fidelity only (facts kept, meaning unchanged, nothing added, greeting/sign-off kept, not a verbatim copy). Chosen = fidelity-clean with the least verbatim overlap; rejected = critical fidelity error or near-copy. No style, length or detector signal enters selection.
• Inference guard: if a sample copies > 35 % of the draft's 5-grams it is resampled once with a logits penalty on draft n-grams (digits exempt).

Evaluation (39-case daily-use set, 2 samples each)

62 English samples	this model	baseline 4B rewriter
samples copying > 35 % of draft	0	33 / 93
critical fidelity errors (judge)	10 %	≈ 30 %
Chinese cases passing judge	10 / 16	4 / 16
Originality.ai "human" (external check only)	79 %	57 %

GGUF files (Q8_0 / Q6_K / bf16; Q5_K_M withheld at 17/62 critical errors, Q4_K_M withheld as gibberish) are in the gguf/ folder of this repo; their quality on the same set is in docs/QUALITY.md of the GitHub repo. MLX 4-/6-bit quantisation of this model is not usable (Gemma 4 PLE layers); use the GGUF quants on a Mac.

Known failure modes: ~1 in 10 outputs has a meaning flip (who did what, ordered vs received, a metric renamed); subject lines/greetings occasionally dropped; Chinese weaker than English; short drafts (< 120 words) less reliable. Proofread numbers, dates and the direction of every claim.

Prompt format (must match exactly)

Base-model continuation, not chat. prompt_format.json in this repo holds the instruction and separator:

{instr}

{draft}

### Rewritten:

Stop at EOS; temperature 0.85, top-p 0.95.

Usage

import json, torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import hf_hub_download

repo = "jialinyyzz/humanizer-gemma-4-e4b"
pf = json.load(open(hf_hub_download(repo, "prompt_format.json")))
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="cuda")
draft = open("draft.txt").read()
prompt = pf["instr"] + "\n\n" + draft.strip() + pf["sep"]
ids = tok(prompt, return_tensors="pt").to("cuda")
out = model.generate(**ids, max_new_tokens=900, do_sample=True, temperature=0.85, top_p=0.95)
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

The full inference path with the anti-copy resample, an MLX server, the evaluation set and the training scripts are on GitHub: sgaofen/humanizer.

Licence

Weights derive from Gemma 4 and are provided under and subject to the Gemma Terms of Use. Code: Apache-2.0.