Paraphase-4B — AI humanizer for email

Paraphase-4B is a 4.2B-parameter AI humanizer for English email. It rewrites AI-sounding drafts so they read like a person wrote them quickly at work — keeps every number, name, date and commitment, adds nothing, and is 100% local — no API, no upload. It is a QLoRA fine-tune of Qwen/Qwen3.5-4B, shipped as GGUF for llama.cpp.

The name is deliberate: paraphrase + phrase. This is not a chatbot; it is a rewriter with a single job.

Task adapters (new)

The raw Qwen3.5-4B base, the rewrite LoRA, and the measured extraction LoRA are in the Paraphase-4B Adapter Pack.

On a frozen 50-email AESLC test, the extraction LoRA raises exact match from 0.12 to 0.42, micro F1 from 0.516 to 0.670, and date F1 from 0.407 to 0.892. The tool path is the raw base plus a narrow guard: 0.70 to 0.75 on the frozen 20-item tool instrument. The tool LoRA did not beat the base and is not shipped.

Important: the paraphase-4b-Q8_0.gguf in this repo is the merged humanizer, not the adapter base. Use the Adapter Pack's raw qwen3.5-4b-Q8_0.gguf with the LoRAs.

The Paraphase collection → — the 4B, 2B, adapter pack, live demo, and explorer, with the premise in one place.

Quick start

1. Download paraphase-4b-Q8_0.gguf (4.48 GB).
2. Serve it: llama-server -m paraphase-4b-Q8_0.gguf -c 2048
3. Send the rewrite prompt with "chat_template_kwargs": {"enable_thinking": false}. Full usage and the retry policy ↓

Try it live → — paste a draft, get a rewrite on ZeroGPU.

Open the interactive comparison → — the before/after, the Pangram data, and the 2B/4B/Q8/Q4 trade-offs. Smaller sibling: Paraphase-2B (2.68 GB Q8; needs the retry wrapper).

Before and after

Three synthetic drafts, rewritten by the Q8 + retry policy. The measured numbers are below; these are demos, not a benchmark.

Quarterly planning

• Before: I hope this email finds you well. I am writing to inform you that the quarterly planning meeting has been rescheduled to Thursday, November 6 at 2:00 PM in Conference Room B. Please ensure that your team's updated projections are submitted by end of day Wednesday.
• After: Hi, just wanted to let you know the quarterly planning meeting is moved to Thursday, Nov 6 at 2:00 PM in Conference Room B. Make sure your team's updated projections are in by end of day Wednesday.

September expenses

• Before: Furthermore, it is important to note that all expense reports for the month of September must be submitted by October 15. Late submissions will not be processed until the following cycle.
• After: Also, all expense reports for September must be submitted by October 15. Late reports will not be processed until the next cycle.

Vendor signatures

• Before: Per my previous message, the vendor agreement requires signatures from both parties by Friday. Kindly advise if there are any outstanding questions.
• After: As discussed earlier, the vendor agreement needs signatures from both sides by Friday. Let me know if you have any remaining questions.

What Pangram measured

Detector claims about "humanizers" are usually unverifiable. This model was tested with a pre-registered Pangram bulk mission, and the result is reported with its limits.

The 186-probe mission. One bulk batch, 186 probes, 397 units, $15.88: 50 human originals, 50 AI-ified drafts, 50 rewrites of those drafts, plus a 12-item human-source gate and a 12-item Q4 check. Human originals passed 50/50. Only 13/50 drafts were flagged at fraction_ai ≥ 0.5, so the drafts were a weak AI-detection challenge and the mission's product-path arm is uninformative by its own control (33/50 drafts carried a non-"Human" Pangram category, but the pre-registered cut was the fraction).

The flagged subset (the case that matters). The 13 drafts Pangram actually flagged were frozen, and the shipped 4B + contract-retry policy was tested on them in a second pre-registered run (13 probes, $0.88):

metric	result
improved	10/13
median fraction_ai change	−0.588 (95% CI [−0.794, 0.000])
exact sign test	p = 0.039
crossed below 0.5 ("rescued")	9/13
pass → fail flips	0

The pre-registered gate was stricter — at least 11/13 improved — so this is suggestive, not a confirmed detector win. The median item and the rescue count moved in the right direction; two drafts got worse and one did not move. A 2B sibling showed the same shape (10/13, median −0.503, 7 rescues).

The 12-item human-source gate. The 4B + retry passed 11/12 vs the base model's 11/12 (one discordant item each way, McNemar p = 1.0; median fraction_ai change −0.210). The 2B sibling passed 12/12 vs 11/12. The pass-rate difference is descriptive only.

What this means: on drafts Pangram flags, the model plus retry lowers Pangram's AI fraction substantially and moves most of them into the human range in this sample. It is not a guarantee of passing as human-written, and the strict pre-registered gate was narrowly missed. Raw scores are in pangram/ (aggregates and per-item scores; no email text).

Measured quality (voice50, 50 paired items; Q8 rows include the retry policy)

model	valid	tells	fact fidelity	copy reuse	length	tool calls
4B Q8 + retry (recommended)	50/50	0.08	0.97	0.26	1.11	0.65
4B Q8 greedy	50/50	0.10	0.97	0.43	1.14	0.65
4B Q4 greedy (experimental)	48/50	0.15	0.97	0.53	0.84	0.45

The retry policy is part of the recommendation: greedy output is already 100% valid, but resampling outputs that copy more than 70% of the input lowers copy reuse from 0.43 to 0.26. Q4 is smaller (2.6 GB) but misses the release gates on copy reuse and tool calls; use Q8 unless VRAM forces Q4.

How to use

llama.cpp

llama-server -m paraphase-4b-Q8_0.gguf -c 2048 --port 8080

Send the rewrite prompt with thinking disabled (Qwen3.5's template otherwise starts with a "Thinking Process" preamble):

import requests

PROMPT = (
    "Rewrite the email below so it reads like a real person wrote it quickly at work. "
    "Keep every fact, number, name, and commitment exactly the same. "
    "Use plain, direct language. No em dashes. Keep it about the same length. "
    "Output only the rewritten email.\n\nEMAIL:\n"
)

resp = requests.post("http://127.0.0.1:8080/v1/chat/completions", json={
    "messages": [{"role": "user", "content": PROMPT + draft}],
    "temperature": 0,
    "max_tokens": 1024,
    "chat_template_kwargs": {"enable_thinking": False},
})
print(resp.json()["choices"][0]["message"]["content"])

Contract-aware retry (recommended)

Take the greedy output. If it is empty/meta/instruction-echo, shorter than max(20, 0.5 × input) words, or reuses more than 70% of the input's 5-grams, resample at temperature 0.9 (at most twice) and keep the first acceptable candidate. No detector is used; the check is input/output-only. This is the policy measured above.

Files

file	size	notes
paraphase-4b-Q8_0.gguf	4.5 GB	recommended rung
paraphase-4b-Q4_K_M.gguf	2.7 GB	experimental; below release gates
adapter/	101 MB	LoRA adapter (r=16, all linear layers)

Training

• Base: Qwen/Qwen3.5-4B (Apache-2.0).
• Method: 4-bit QLoRA, 200 steps, LR 1e-4, r=16 on q/k/v/o/gate/up/down.
• Data mix: 502 low-overlap paraphrase pairs (teacher Qwen3.5-4B, source-to-target 5-gram reuse mean 0.094) ×2, 175 notes→email pairs, 475 real AI-draft/human pairs, 412 persona rows, 26 tool-replay rows ×2.
• Human source text is AESLC-derived and is not redistributed here.
• No detector was used in training.

Limitations

• English email only; not for documents, chat, or non-English text.
• The detector evidence is one vendor (Pangram 4.1), one run per probe, and a suggestive flagged-subset result; no claim of universal detector evasion.
• 4B Q8 is 4.5 GB; a 6 GB GPU is the practical floor.
• The model can still copy heavily on some inputs; the retry policy is part of the recommended setup.
• No human prose-quality evaluation has been run.
• Q4 is included for small machines but does not meet the Q8 release gates.

License

Apache-2.0, inherited from the base model. See LICENSE.

Citation

@misc{paraphase4b,
  title  = {Paraphase-4B: a local humanizer and paraphraser for email},
  author = {Chad St. Thomas},
  year   = {2026},
  note   = {QLoRA fine-tune of Qwen3.5-4B; Pangram bulk evaluation}
}