sriyuth/ari-rewritebot-135m-dpo

Ari-LLM RewriteBot 135M (DPO-Aligned)

Ari-LLM RewriteBot (135M) is an ultra-fast, commercial-grade paraphrasing and text rewriting foundation model developed by Ari-LLM.

It is trained on the curated 10-Mode QuillBot & PAWS & QQP corpus with full-parameter SFT, followed by Direct Preference Optimization (DPO) to punish lazy verbatim repetition and reward high lexical diversity, clause restructuring, and stylistic fidelity.

Designed specifically to run 100% locally on consumer CPUs with ultra-low memory footprint (~540MB) and instantaneous latency.

---

Supported Modes & Control Tokens

Ari-LLM RewriteBot supports 10 distinct rewriting modes via system prefix tags:

Mode Tag	Purpose	Example Transformation
[Standard Paraphrase]	Natural clause and synonym reformulation	Severe weather canceled the flight ➔ The flight was canceled due to bad weather.
[Formal Rewrite]	Professional email & business correspondence	Send me that soon ➔ Could you kindly provide the requested document at your earliest convenience?
[Academic Rewrite]	Scholarly, research-paper phrasing	This shows it's true ➔ These empirical findings substantiate the hypothesis across test observations.
[Simple Rewrite]	Plain English & maximum clarity	Utilizing methodologies facilitates efficiencies ➔ Using modern computers makes work faster and easier.
[Creative Rewrite]	Vivid metaphors & literary prose	The sun went down ➔ Twilight melted into twilight shadows as diamonds pierced the velvet expanse of night.
[Shorten Mode]	Executive summary & high conciseness	Due to the circumstance that... ➔ Because of...
[Anti-AI Humanize]	Removes robotic AI syntax & clichés	In conclusion, it is important to remember... ➔ Look, navigating daily life gets messy—and that's totally normal.
[Active Voice]	Restructures passive sentences to active	The bug was found by QA ➔ The QA team discovered the problem.
[Grammar Polish]	Corrects syntax and grammatical errors	She don't know the answer ➔ She does not know the answer.
[Flip Clauses & Invert]	Inverts subordinate clauses for fresh flow	Reverses cause-effect sentence architecture

---

Quickstart (Python & Transformers)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "sriyuthlucky/ari-rewritebot-135m-dpo" # or local path
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, dtype=torch.float32)

def rewrite(text: str, mode: str = "Standard Paraphrase") -> str:
    prompt = f"<|im_start|>user\n[{mode}] {text}<|im_end|>\n<|im_start|>assistant\n"
    inputs = tokenizer(prompt, return_tensors="pt")
    
    with torch.no_grad():
        out = model.generate(
            **inputs,
            max_new_tokens=100,
            do_sample=True,
            temperature=0.7,
            top_p=0.9,
            repetition_penalty=1.15,
            eos_token_id=[tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<|im_end|>")],
            pad_token_id=tokenizer.eos_token_id
        )
    gen = out[0][inputs["input_ids"].shape[1]:]
    res = tokenizer.decode(gen, skip_special_tokens=True).strip()
    return res.split("<|im_end|>")[0].strip()

# Example:
text = "The flight was canceled because of severe weather conditions."
print(rewrite(text, mode="Standard Paraphrase"))
# Output: "The flight was canceled due to bad weather."

print(rewrite("Can you send me that project update as soon as possible?", mode="Formal Rewrite"))
# Output: "Could you kindly provide the requested project update at your earliest convenience?"

---

Training Methodology

1. Foundation Base: SmolLM2-135M pre-trained on 2 Trillion high-quality English tokens.
2. Supervised Fine-Tuning (SFT): Full-parameter fine-tuning on 113,000 multi-mode paraphrase pairs (PAWS + QQP + QuillBot master templates) with prompt loss masking (-100).
3. Direct Preference Optimization (DPO):
   • Reference Model: Frozen SFT Base
   • Chosen ($y_w$): Lexically rich, structurally inverted rewrites
   • Rejected ($y_l$): Verbatim lazy copies, repetitive phrasing
   • Reward Margin reached: +3.30

---

Citation & Branding

@misc{ari_rewritebot_2026,
  author = {Ari-LLM Team},
  title = {Ari-LLM RewriteBot 135M: High-Performance DPO Paraphraser},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/sriyuthlucky/ari-rewritebot-135m-dpo}}
}