humanization-by-iterative-paraphrasing

KO | EN

gitlite — search

#codex #api #automation #java #ai #rag #containers #data-science #privacy #c

humanization-by-iterative-paraphrasing

(★ 9)

No description available.

File Explorer

ZIP Download

- - eval_gptzero_smoke.yaml
  - eval_pangram_smoke.yaml
  - eval_semantic_openai_smoke.yaml
  - infer_qwen3_0_6b_smoke.yaml
  - train_qwen3_0_6b_smoke.yaml
- deepspeed_zero2_qlora.json
- eval_gptzero.yaml
- eval_pangram.yaml
- eval_semantic_openai.yaml
- infer_llama3_70b_hip.yaml
- infer_llama3_8b_hip.yaml
- infer_qwen3_0_6b_hip.yaml
- train_llama3_70b_hip_qlora.yaml
- train_llama3_8b_hip.yaml
- train_qwen3_0_6b_hip.yaml
- input_texts.jsonl
- .gitkeep
- .gitkeep
- smoke_test.sh
- - __init__.py
  - config.py
  - data.py
  - evaluate.py
  - inference.py
  - prompts.py
  - train.py
.gitignore
LICENSE
pyproject.toml
README.md
uv.lock

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

This repository contains the training and inference code for Humanization by Iterative Paraphrasing (HIP). HIP trains a LoRA adapter to rewrite an AI paraphrase back toward human-written prose, then applies that paraphraser iteratively at inference time.

The repository provides:

- HIP LoRA training on AI-to-human rewrite pairs.
- Iterative HIP inference with trained or released adapters.
- Evaluation utilities matching the paper's GPT-5 nano semantic judge, GPTZero, and Pangram scoring.
- Example configs for small-model smoke tests, representative 8B training, and 70B QLoRA training.

For evaluation, provide API keys through environment variables such as `OPENAI_API_KEY`, `GPTZERO_API_KEY`, and `PANGRAM_API_KEY`.

## Repository Layout

```text
configs/                 Example train, inference, and scoring configs
configs/smoke/           Small smoke-test configs
examples/                Tiny JSONL input for smoke inference
scripts/smoke_test.sh    Convenience script for the smoke test
src/hip/train.py         LoRA/QLoRA training entry point
src/hip/inference.py     Iterative HIP inference entry point
src/hip/evaluate.py      OpenAI semantic, GPTZero, and Pangram evaluation entry point
outputs/                 Ignored output directory
inputs/                  Ignored local input directory
```

## Artifacts

The companion Hugging Face [collection](https://huggingface.co/collections/YixuanEvenXu/humanization-by-iterative-paraphrasing) contains the released adapters and data.

The dataset exposes two configs:

```python
from datasets import load_dataset

training = load_dataset("YixuanEvenXu/HIP-training-and-evaluation-data", "training")
prefix_eval = load_dataset("YixuanEvenXu/HIP-training-and-evaluation-data", "prefix_eval")
```

For training, use `paraphrased_text` as the HIP input and `text` as the target.

## Installation

```bash
uv sync
```

For the advanced 70B QLoRA recipe, also install the optional quantization dependency:

```bash
uv sync --extra qlora
```

You also need Hugging Face access for gated base models such as Llama-3.

## Quick Smoke Test

The smoke test uses `Qwen/Qwen3-0.6B-Base` and runs one training step on eight examples. It verifies that the training, inference, and API evaluation paths are wired correctly.

```bash
uv run hip-train --config configs/smoke/train_qwen3_0_6b_smoke.yaml
uv run hip-run --config configs/smoke/infer_qwen3_0_6b_smoke.yaml
OPENAI_API_KEY=... uv run hip-score --config configs/smoke/eval_semantic_openai_smoke.yaml
GPTZERO_API_KEY=... uv run hip-score --config configs/smoke/eval_gptzero_smoke.yaml
PANGRAM_API_KEY=... uv run hip-score --config configs/smoke/eval_pangram_smoke.yaml
```

Outputs are written under `outputs/smoke/qwen3_0_6b/`.

## Hardware Expectations

Paper-scale training used:

- Most adapters: `4 x NVIDIA L40S`.
- Llama-3-70B adapters: `8 x NVIDIA L40S` with QLoRA-style loading.

These are paper-run resources, not strict minimums. Smaller models can often be trained on fewer GPUs by reducing batch size, increasing gradient accumulation, shortening sequence length, or using quantized loading. The 70B recipe is the least portable and may require environment-specific adjustment.

## Training

A full small-model run uses the released training split:

```bash
uv run hip-train --config configs/train_qwen3_0_6b_hip.yaml
```

The Llama-3-8B training recipe is:

```bash
accelerate launch --num_processes 4 \
  -m hip.train --config configs/train_llama3_8b_hip.yaml
```

The Llama-3-70B QLoRA recipe is more hardware- and environment-sensitive:

```bash
uv sync --extra qlora
accelerate launch --num_processes 8 \
  -m hip.train --config configs/train_llama3_70b_hip_qlora.yaml
```

## Inference

Run iterative HIP with a released adapter from Hugging Face:

```bash
uv run hip-run --config configs/infer_qwen3_0_6b_hip.yaml
uv run hip-run --config configs/infer_llama3_8b_hip.yaml
uv run hip-run --config configs/infer_llama3_70b_hip.yaml
```

The input file can be a JSONL, parquet, or Hugging Face dataset split. Each row should contain the configured `text_field`. The output parquet contains one row per source example per HIP round with:

- `original_text`
- `input_text`
- `output_text`
- `round`
- `original_word_count`
- `input_word_count`
- `output_word_count`

## Evaluation

`hip-score` scores an inference parquet with one of three providers selected by the config:

- `openai_semantic`: GPT-5 nano semantic-preservation judge. Adds `gpt5_nano_semantic_score`.
- `gptzero`: GPTZero detector API. Adds GPTZero class probabilities, including `gptzero_prob_human`.
- `pangram`: Pangram detector API. Adds Pangram fractions, including `pangram_fraction_human`.

Configs read API keys from `OPENAI_API_KEY`, `GPTZERO_API_KEY`, or `PANGRAM_API_KEY`. You can also set `api_key` locally in a config for a local run; environment variables are preferred for shared configs.

```bash
uv run hip-score --config configs/eval_semantic_openai.yaml
uv run hip-score --config configs/eval_gptzero.yaml
uv run hip-score --config configs/eval_pangram.yaml
```

Detector access, pricing, and rate limits are provider-dependent.

## Scope and Safety

This code is released for research reproducibility and detector robustness analysis. It should not be used to deploy systems for evading detection in settings where that violates policy, law, or platform rules.

Keep API keys, detector credentials, trained checkpoints, and generated evaluation outputs outside version control.

## Citation

```bibtex
@article{xu2026base,
  title={Base Models Look Human To AI Detectors},
  author={Yixuan Even Xu and Ziqian Zhong and Aditi Raghunathan and Fei Fang and J. Zico Kolter},
  journal={arXiv preprint arXiv:2605.19516},
  year={2026}
}
```

humanization-by-iterative-paraphrasing

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

humanization-by-iterative-paraphrasing

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README