Rusa banner with the words "Rusa. Better Human."

Rusa-1

Rusa-1 is a 350M-parameter model that rewrites AI-sounding text so it reads like a person wrote it, while keeping every fact, number, name and quote intact.

It is small enough to run on a CPU, on a phone, or fully in the browser.

  • Base model: LiquidAI/LFM2-350M
  • Training: distilled from a 12B rewriting model, then preference-tuned on its own samples, with a 12B teacher ranking them and a fact checker filtering them
  • Language: English

V1.1 files

The standard V1.1 GGUFs are in gguf/. The experimental ternary builds are available separately. Rusa-1-Small

Format Path Size Use with
Safetensors (BF16) model.safetensors ~710 MB πŸ€— Transformers, vLLM
GGUF gguf/Rusa-{F16,Q8_0,Q6_K,Q5_K_M,Q4_K_M}.gguf Q4_K_M 229.31 MB, Q5_K_M 260.37 MB, Q6_K 293.38 MB, Q8_0 379.21 MB, F16 711.48 MB llama.cpp, LM Studio, Ollama
ONNX onnx/model{,_fp16,_quantized,_q4,_q4f16}.onnx q4f16 248 MB, q4 289 MB, q8 496 MB, fp16 692 MB, fp32 1.35 GB Transformers.js v4 (browser), ONNX Runtime

Q8_0 is the recommended GGUF. At 350M parameters, 4-bit quantization costs noticeably more quality than it does on large models.

Prompt format

Rusa uses a plain-text prompt, not a chat template. Put the draft between the instruction and the ### Rewritten: marker. The model stops by itself with <|im_end|>.

Rewrite the text below so it reads like a person wrote it, not a language model.

Reorganize it as you see fit. Vary sentence length on purpose. Cut hedging,
throat-clearing, and any sentence that only announces what comes next.
Prefer the concrete word over the abstract one. It is fine to sound uneven.

Every fact, number, unit, date, name and quotation must survive unchanged.

{your text}

### Rewritten:

Recommended sampling: temperature=1.0, top_p=0.95, top_k=0 (disabled), no repetition penalty. Sampling a few rewrites and keeping the best one works well. With 8 samples, at least one keeps every fact on 89% of the test drafts.

Usage

Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM

INSTR = """Rewrite the text below so it reads like a person wrote it, not a language model.

Reorganize it as you see fit. Vary sentence length on purpose. Cut hedging,
throat-clearing, and any sentence that only announces what comes next.
Prefer the concrete word over the abstract one. It is fine to sound uneven.

Every fact, number, unit, date, name and quotation must survive unchanged."""

repo = "jun-ai/Rusa-1"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto")

draft = "In today's fast-paced world, effective time management is more important than ever."
inputs = tok(f"{INSTR}\n\n{draft}\n\n### Rewritten:\n\n", return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=1.0, top_p=0.95, top_k=0)
print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

llama.cpp

llama-cli -hf jun-ai/Rusa-1:Q8_0 --temp 1.0 --top-p 0.95 --top-k 0 -f prompt.txt

prompt.txt holds the full prompt shown above.

Transformers.js (browser)

import { pipeline } from "@huggingface/transformers";

const rewrite = await pipeline("text-generation", "jun-ai/Rusa-1", { dtype: "q4f16", device: "webgpu" });
const prompt = `${INSTR}\n\n${draft}\n\n### Rewritten:\n\n`;
const [out] = await rewrite(prompt, {
  max_new_tokens: 1024, do_sample: true, temperature: 1.0, top_p: 0.95, top_k: 0, return_full_text: false,
});
console.log(out.generated_text);

Use dtype: "q8" or "fp16" for higher quality, or device: "wasm" where WebGPU is not available.

Evaluation

The test set holds 209 held-out AI-written drafts across essays, emails, reviews, forum posts and how-tos. Each draft got one sampled rewrite. AI likelihood comes from the desklib/ai-text-detector-v1.01 classifier (lower is better). A rewrite counts as flagged when its score is above 0.5. A rewrite "keeps the facts" when every number, date, name and quote from the draft survives.

Model Size Mean AI score ↓ Flagged as AI ↓ Keeps all facts ↑ Keeps all facts, best of 8 ↑
12B teacher 12B 0.190 30 / 209 67 / 209 146 / 209
Rusa 350M 0.236 46 / 209 83 / 209 185 / 209

Rusa is 35Γ— smaller than its teacher and keeps facts more reliably. Its detector score is slightly higher than the teacher's.

Limitations

  • English only.
  • Fact keeping is good, not perfect. Check numbers, names and quotes before you publish a rewrite.
  • Long inputs (more than ~450 words) are outside the training range. Split long documents into paragraphs.
  • Do not use Rusa to misrepresent authorship where that is prohibited, such as in academic work.

License

Rusa inherits the LFM Open License v1.0 from LFM2-350M.

Downloads last month
129
Safetensors
Model size
0.4B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for jun-ai/Rusa-1

Quantized
(44)
this model
Quantizations
1 model