humanizer-1B-OptIQ-4bit

A 1B model that scores the same as the human reference set on the RADAR AI detector. Stacked SFT + DPO LoRA adapters on top of mlx-community/MiniCPM5-1B-OptIQ-4bit close 100% of the gap to human writing on a 200-draft held-out evaluation.

	P(AI) (RADAR-Vicuna-7B)
Source AI drafts (Qwen3.5-4B + Gemma-4-e4b output)	0.51
humanizer-1B-OptIQ-4bit (SFT + DPO stacked)	0.37
Human reference (EditLens ICLR 2026, n=200)	0.37

Build, recipe, and discussion: https://mlx-optiq.com/blog/humanizer-stacked-lora

What's in this repo

humanizer-1B-OptIQ-4bit/
  model.safetensors, config.json, tokenizer*    base MiniCPM5-1B-OptIQ-4bit
  optiq_metadata.json                            per-layer bit assignments
  adapters/
    humanizer-sft/                               SFT humanizer LoRA
      adapters.safetensors
      adapter_config.json
      optiq_lora_config.json
    humanizer-dpo/                               DPO continuation LoRA
      adapters.safetensors
      adapter_config.json
      optiq_lora_config.json

• Base. mlx-community/MiniCPM5-1B-OptiQ-4bit. OptIQ mixed-precision quant of openbmb/MiniCPM5-1B. 875 MB on disk, Capability Score 30.28.
• SFT adapter. Trained on canonical SFT data derived from the EditLens ICLR 2026 corpus. --preset large (ranks 32 and 64, with the by_bits overlay), 600 iters, mask_prompt=True.
• DPO adapter. Trained as a delta on top of the SFT via optiq lora train --method dpo --mount-adapter. The reference KL is anchored against base + SFT (the textbook SFT then DPO continuation), so the saved adapter contains only the DPO delta. 300 iters, beta 0.1, LR 5e-5 with linear warmup then cosine decay (the OptIQ DPO defaults).

The DPO adapter is meaningful only when applied alongside the SFT adapter. It is a delta from the SFT distribution, not a standalone LoRA. Apply both at inference for the headline result.

Use

You need mlx-optiq >= 0.1.4 for the multi-LoRA serving and stacking syntax:

pip install 'mlx-optiq>=0.1.4'

# Download the repo
huggingface-cli download mlx-community/humanizer-1B-OptIQ-4bit \
  --local-dir ./humanizer-1B-OptIQ-4bit

# Serve with both adapters mounted
optiq serve \
  --model ./humanizer-1B-OptIQ-4bit \
  --adapter ./humanizer-1B-OptIQ-4bit/adapters/humanizer-sft \
  --adapter ./humanizer-1B-OptIQ-4bit/adapters/humanizer-dpo \
  --port 8080

Send requests with both adapters active via the + stacking syntax in the request body:

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "./humanizer-1B-OptIQ-4bit",
    "adapter": "humanizer-sft+humanizer-dpo",
    "messages": [
      {"role": "system", "content": "Rewrite AI-generated drafts into natural human-style prose, preserving meaning, facts, names, numbers, citations, URLs, quotes, and formatting."},
      {"role": "user", "content": "STYLE: direct technical blog\nTONE: analytical, clear, non-corporate\nLENGTH: preserve within 15%\n\nDraft to rewrite:\n\n[your AI-generated draft here]"}
    ],
    "temperature": 0.4,
    "max_tokens": 1600,
    "chat_template_kwargs": {"enable_thinking": false}
  }'

The OpenAI-compatible endpoint is a drop-in for Open WebUI, Continue, Cursor, your own scripts. Send "adapter": "humanizer-sft" to use SFT alone, or "adapter": "base" to bypass adapters entirely (useful for A/B comparisons).

Held-out evaluation

200 AI-generated drafts from the EditLens ICLR 2026 held-out set, rewritten by each system and scored by RADAR-Vicuna-7B. Lower P(AI) is more human-like.

Pipeline	P(AI)	Delta vs source	Slop / 1K tokens
Source AI draft (Qwen3.5-4B + Gemma-4-e4b)	0.51	—	0.6
SFT humanizer alone	0.50	-0.01	0.2
SFT + DPO stacked (this repo)	0.37	-0.14	0.0
Human reference (target)	0.37	-0.14	0.1

The stacked pipeline produces fewer slop phrases per 1K tokens (0.0) than the human reference set itself (0.1).

Intended use and limitations

• Intended use. Rewriting AI-generated drafts (blog posts, articles, reports) into more natural-sounding prose. Preserves facts, names, numbers, URLs, citations.
• Trained on. The EditLens ICLR 2026 corpus filtered through the OptIQ Labs dataset-building pipeline. Qwen3.5-4B and Gemma-4-e4b were the source AI models, the original EditLens human-written prose was the target.
• AI-detector caveat. RADAR-Vicuna-7B is one detector out of many. Matching the human reference on RADAR means the rewrites land at the same point on RADAR's scale as the EditLens human-written set. Other detectors will give different numbers, and detector arms races mean any specific score has a shelf life. The reproducible claim is the delta from source and the gap closure against a fixed human reference. Both held up across the entire 200-draft held-out set.
• Length. The rewrites tend to over-generate (length ratio around 3 to 4 times the source). Apply a max-tokens or post-truncation step if you need length-faithful output.
• Capability outside humanization. This LoRA stack is heavily specialized for the rewrite-this-AI-draft format. Out-of-format prompts will degrade behavior. Serve "adapter": "base" for general MiniCPM5-1B inference.

License

• Base model: openbmb/MiniCPM5-1B (Apache-2.0).
• LoRA adapters: Apache-2.0, this release.
• Training data: derived from EditLens ICLR 2026 (research use).