cive202/humanize-ai-text-mistral-7b-lora

LoRA/QLoRA adapter for AI → Human text rewriting (“humanization”) on top of mistralai/Mistral-7B-Instruct-v0.2.

• What this repo contains: adapter weights only (PEFT/LoRA), not the full base model
• What it does: rewrites AI-styled passages into more human-like writing while preserving meaning
• How to use: load the base model + attach this adapter via peft

---

What “AI → Human humanization” means here

Given an AI-generated passage (x), the model generates a rewrite (\hat{y}) that aims to satisfy:

• Semantic preservation (retain meaning)
• Stylistic transformation (shift measurable markers toward human distributions)

This repo is part of a broader comparison:

• Encoder–decoder (BART)
• Decoder-only LLM + adapters (Mistral)

---

📄 Paper

“Rewriting the Machine: Encoder-Decoder vs. Decoder-Only Transformers for AI-to-Human Text Style Transfer”
Authors: Utsav Paneru et al.
arXiv: https://arxiv.org/abs/2604.11687v1
Status: Preprint (2026)

Citation

@misc{paneru2026makesoundlikehuman,
      title={Please Make it Sound like Human: Encoder-Decoder vs. Decoder-Only Transformers for AI-to-Human Text Style Transfer}, 
      author={Utsav Paneru},
      year={2026},
      eprint={2604.11687},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.11687}, 
}

Quickstart (Transformers + PEFT)

Mistral is a gated model. You may need to accept terms and run huggingface-cli login.

pip install -U "transformers>=4.40.0" "peft>=0.10.0" accelerate bitsandbytes torch

Load base + adapter (recommended: 4-bit inference)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_id = "mistralai/Mistral-7B-Instruct-v0.2"
adapter_id = "cive202/humanize-ai-text-mistral-7b-lora"

tokenizer = AutoTokenizer.from_pretrained(base_id)

base = AutoModelForCausalLM.from_pretrained(
    base_id,
    device_map="auto",
    torch_dtype=torch.float16,
    load_in_4bit=True,
)

model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

ai_text = "Large language models often produce fluent, structured prose with recognizable regularities..."

prompt = f"""### Instruction:
Rewrite the input so it sounds human-written while preserving meaning.

### Input:
{ai_text}

### Response:
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=True,
        temperature=0.7,
        top_p=0.9,
        repetition_penalty=1.1,
        eos_token_id=tokenizer.eos_token_id,
    )

print(tokenizer.decode(out[0], skip_special_tokens=True))

---

Training summary (adapter config)

• Base model: mistralai/Mistral-7B-Instruct-v0.2
• Quantization: 4-bit NF4, double quantization, float16 compute
• LoRA: r = 16, α = 32, dropout = 0.05
• Target modules: q_proj, k_proj, v_proj, o_proj
• Max sequence length: 512
• Optimizer: Paged AdamW 32-bit
• LR / scheduler: 2e-4, cosine, warmup ratio 0.05
• Training length: 500 steps (checkpoint every 100)
• Effective batch size: 8
• Loss: completion-only (masked before ### Response:)

---

Dataset

Parallel chunk pairs via sentence-aware chunking:

• Train: 25,140 pairs
• Validation: 1,390
• Test: 1,390

Preprocessing

• Sentence tokenization (NLTK)
• Greedy packing (≤200 tokens)
• Remove short pairs (<10 words)
• Document-disjoint splits

Metadata fields: doc_id, chunk_idx, ai, human, style, model, prompt_id.

---

Evaluation (test n = 1,390)

Reference similarity

• BERTScore F1: 0.8980
• ROUGE-L: 0.4642
• chrF++: 55.6770

Fluency proxy

• GPT-2 PPL (output): 9.0325
• GPT-2 PPL (AI input): 37.8485
• GPT-2 PPL (human): 23.6912

Note: Very low perplexity may indicate overly predictable text, not necessarily human-like.

Style shift

• Mean marker shift: 1.2788

Highlights:

• Overshooting on multiple markers (capped at 2.0)
• Comma usage shifts in the wrong direction (-1.0)

---

Limitations

• Overshooting human style distributions
• Some features move in the wrong direction
• Not guaranteed to bypass AI detectors
• Performance depends on domain similarity

---

Reproducibility

Training + evaluation pipeline comes from the unpublished BARTvsMistral project.

Evaluation outputs (e.g., summary.json) can be used to regenerate figures externally.

---

License

• Adapter weights: Apache-2.0
• Base model: follow mistralai/Mistral-7B-Instruct-v0.2 license terms