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.

v2 (2026-09-14): SFT + DPO + GRPO. Trained on data that teaches the model how people write, then reinforced on fidelity only. No detector was used as a reward, filter, signal or selection criterion at any stage; 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.
• GRPO, 300 steps (v2). LoRA on the merged SFT + DPO model, 8 samples per draft. Reward = LLM-judged fidelity against an atomic fact list of the draft (critical −3.0, minor −0.15, invented −2.0, reversed −2.0, dropped format element / greeting −1.5), −1.0 if the draft's register is normalised, and a superlinear reuse ramp (reuse = max(verbatim 5-gram copy, content-masked syntactic-skeleton recall); free below 0.31, −4 at full copy). Errors are credited to the sentence that carries them. Release checkpoint picked from a 25-step ladder on the held-out set. No detector anywhere in the reward.
• 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)

Two independent judges (GLM-5.3 and gpt-5.6); an error counts only if both report it. v1 re-graded under the same protocol.

62 English samples	v2	v1 (SFT + DPO)	baseline 4B rewriter
samples copying > 35 % of draft	0	1	33 / 93
reuse (verbatim ∨ syntactic skeleton), median	0.29	0.34	—
critical fidelity errors	0 / 62	3 / 62	≈ 30 %
minor / clean	20 / 42	15 / 44	—
format element dropped	5 / 62	12 / 62	—
Chinese cases passing judge	13 / 16	11 / 16	4 / 16
Originality.ai "human" (external check only)	85 %	81 %	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 (v2): dropped qualifiers in about 1 in 3 outputs ("an estimated 4.2 %" → "4.2 %", "suggest" → "conclude"); meaning flips were v1's main failure (~1 in 10) and did not appear in v2's 62 samples, but the set is small; an occasional garbled sentence — resample; Chinese weaker than English (13/16); short drafts (< 120 words) less reliable. Proofread numbers, dates and the direction of every claim. GGUF fidelity numbers in docs/QUALITY.md were measured on v1; v2 quants were built with the same recipe.

Prompt format (base-model completion — NOT a chat model)

No system prompt, no chat template, no turn markers. Hand it this exact text and let it continue:

Rewrite the text below so it reads like a person wrote it, not a language model.

Reorganize it as you see fit. Vary sentence length on purpose. Cut hedging,
throat-clearing, and any sentence that only announces what comes next.
Prefer the concrete word over the abstract one. It is fine to sound uneven.

Every fact, number, unit, date, name and quotation must survive unchanged.

<YOUR DRAFT GOES HERE>

### Rewritten:

Instruction and separator ship verbatim in prompt_format.json in this repo (instr, sep; the separator is literally \n\n### Rewritten:\n\n — keep the blank line after it). Reproduce byte for byte; the model was trained on this wrapper and a paraphrased instruction measurably degrades it.

Stop at EOS; temperature 0.85, top-p 0.95, ~900 new tokens.

Ollama / LM Studio apply the base model's chat template by default and that breaks the format. Turn it off — for Ollama, a Modelfile that passes the prompt through untouched:

FROM ./humanizer-gemma-4-e4b-Q8_0.gguf
TEMPLATE """{{ .Prompt }}"""
PARAMETER temperature 0.85
PARAMETER top_p 0.95
PARAMETER num_predict 900

then send the whole block above as one prompt. Our own llama.cpp path is humanizer/gguf_infer.py on GitHub (builds the prompt from prompt_format.json, plus the anti-copy guard); the Modelfile is the hand-rolled equivalent and is untested by us.

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))

Version history: v2 (2026-09-14, this revision) adds the GRPO stage; v1 (2026-09-10, SFT + DPO) is in the commit history of this repo. 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.