qwen2.5-14b-humanizer-v3-lora

🛡️ Qwen2.5-14B Humanizer v3 (Anti-AI Fact-Locked LoRA)

A state-of-the-art 14-Billion parameter Anti-AI Humanizer fine-tuned on top of Qwen/Qwen2.5-14B-Instruct. Engineered specifically to transform formulaic, robotic AI-generated text into organic, engaging, human-written prose while strictly preserving 100% of facts, metrics, and figures.

🌟 Key Architecture & Training Highlights

• Base Model: Qwen/Qwen2.5-14B-Instruct
• Adapter Type: LoRA (Rank r=16, lora_alpha=32, lora_dropout=0.05)
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Phase 1 (SFT): 2,850 gold macro-rewrite pairs with ChatML formatting (1,997 strict number-locked + 853 multi-domain general pairs).
• Phase 2 (DPO): 2,850 preference triplets optimized using Direct Preference Optimization to penalize AI clichés and reward authentic human cadence.

🛡️ Dual Quality Filter System

1. Wikipedia 35 Signs Rulebook: Banishes academic/encyclopedic tells—zero em-dashes (—), zero robotic transitions ("Furthermore", "Moreover"), zero trailing participle clauses (", thereby enhancing..."), zero "stands as a testament".
2. Conor Bronsdon AI Rulebook: Banishes corporate/web AI buzzwords—zero Tier 1A markers ("delve", "tapestry", "watershed moment", "game-changer", "poised to", "leverage", "seamless", "unpack"), zero Tier 1B bloat ("utilize", "in order to", "serves as").
3. 100% Strict Metric Lock: Curated pairs guarantee that all percentages (18.4%), currencies ($4.2M), dates, and metrics are preserved faithfully without hallucination.

---

🚀 Quickstart: Load with Transformers & PEFT

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

BASE_MODEL = "Qwen/Qwen2.5-14B-Instruct"
ADAPTER_REPO = "arunsingh80475/qwen2.5-14b-humanizer-v3-lora"

# 1. 4-Bit NF4 Quantization for fast inference (~9-10 GB VRAM)
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
)

# 2. Load Base Model + Tokenizer
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)

# 3. Load v3 Humanizer Adapter
model = PeftModel.from_pretrained(base_model, ADAPTER_REPO)
model.eval()

# 4. Humanize Text
system_prompt = (
    "You are an authentic, thoughtful human writer. Rewrite the input text into natural, clear, and engaging human prose. "
    "Preserve all core concepts, facts, numbers, and original meaning faithfully, while completely replacing formulaic "
    "AI phrasing, robotic transitions ('Furthermore', 'Moreover'), and repetitive sentence structures with organic human cadence."
)

text_to_humanize = "Our Q3 revenue increased by 18.4% to $4.2 million, serving as a testament to our team's resilience. Furthermore, churn decreased from 6.2% to 3.8%."
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": text_to_humanize}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=400,
        temperature=0.72,
        top_p=0.92,
        repetition_penalty=1.12,
        do_sample=True,
    )

output_text = tokenizer.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True).strip()
print("Humanized Output:
", output_text)

---

💻 Recommended Inference Hyperparameters

---

📜 Citation & Credits

• Author: Arun Singh
• Dataset Pipeline: Multi-Domain Anti-AI Fact-Preservation Benchmark (Wikipedia 35 Signs + Conor Bronsdon Rules)
• Base Model: Qwen Team (Qwen/Qwen2.5-14B-Instruct)