WordSaladChopper

WordSaladChopper (WSC) is a lightweight plug-and-play module that detects and removes “word salad” repetitions in large reasoning models.

---

News

• [11/01/2025] v1 released — first public version now available.

🚀 1. Quick Start

Here we show the on-the-fly (end-to-end) version implemented with Hugging Face. For large-scale paper reproduction, see reproduced/README.md.

Installation

git clone https://github.com/wenyaxie023/WordSaladChopper.git
cd WordSaladChopper
conda create -n wsc python=3.10
conda activate wsc
pip install -e .

Example

This example uses DeepSeek-R1-Distill-Qwen-7B with a ready-to-use classifier hosted on Hugging Face.

from transformers import AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import hf_hub_download
from wscgen.chopper import Chopper
from wscgen.generate import wsc_generate
from wscgen.prober import build_prober
from wscgen.utils import find_newline_token_ids, set_seed

set_seed(41)
model_name = "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Load the trained prober
prober_file = hf_hub_download(
    repo_id="xiewenya/WordSaladChopper_Classifier",
    filename="DeepSeek-R1-Distill-Qwen-7B_s1/probe.pkl",
    repo_type="model",
)
prober = build_prober("logistic").load(prober_file)

# Initialize chopper
chopper = Chopper(
    tokenizer=tokenizer, detector=prober,
    thresh=0.5, streak_len=2, short_streak_len=5, len_threshold=10
)

question = "Return your final response within \\boxed{}. Compute: $1-2+3-4+5- \\dots +99-100$."
messages = [
    {"role": "user", "content": question}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

newline_token_ids = find_newline_token_ids(tokenizer)
gen_cfg = {"temperature": 0.6, "top_p": 0.95}

result_wsc = wsc_generate(
    model, tokenizer, prompt, chopper,
    newline_token_ids=newline_token_ids, gen_cfg=gen_cfg,
    rescue_prompt="I can find a clearer solution if I focus on the core problem.",
    token_budget=32768, rescue_budget=4096, max_rescues=1
)

print("Generated text:", result_wsc["response"])
print("Total tokens used:", result_wsc["total_used_tokens"])

---

2. CLI Usage

2.1 Prepare the Prober

We provide 3 ready-to-use classifiers:

• DeepSeek-R1-Distill-Qwen-7B → DeepSeek-R1-Distill-Qwen-7B_s1/probe.pkl
• DeepSeek-R1-Distill-Qwen-1.5B → DeepSeek-R1-Distill-Qwen-1.5B_s1/probe.pkl
• DeepSeek-R1-Distill-Llama-8B → DeepSeek-R1-Distill-Llama-8B_s1/probe.pkl

To download a probe (example for the DeepSeek-R1-Distill-Qwen-7B):

mkdir -p prober/DeepSeek-R1-Distill-Qwen-7B_s1
wget https://huggingface.co/xiewenya/WordSaladChopper_Classifier/resolve/main/DeepSeek-R1-Distill-Qwen-7B_s1/probe.pkl \
  -O prober/DeepSeek-R1-Distill-Qwen-7B_s1/probe.pkl

2.2 Generate

bash sh/generate.sh

To use your own classifier, set PROBER_PATH in sh/generate.sh before running.

---

3. Training a Custom Classifier (Probe)

The classifier detects hidden-state signals of degeneracy. You can train your own classifier on:

• your own reasoning traces, or
• our released datasets on Hugging Face.

Available Training Data

We provide s1 reasoning traces on Hugging Face for multiple models. You can download them and put them in your data path.

mkdir -p data/DeepSeek-R1-Distill-Qwen-7B_s1
wget https://huggingface.co/datasets/xiewenya/WordSaladChopper_Classifier_Data/resolve/main/DeepSeek-R1-Distill-Qwen-7B_s1/results.json \
     -O data/DeepSeek-R1-Distill-Qwen-7B_s1/results.json

Edit Config

Modify data path in configs/train.yaml before training:

model_name: "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B"

train_mix:
  s1_temp_0_6_top0p95:
    ratio: 1
    temp: 0.6
    top_p: 0.95
    dataset_id: s1_temp_0_6_top0p95
    path: data/DeepSeek-R1-Distill-Qwen-7B_s1/results.json

Start Training

bash sh/training.sh

Using Your Own Dataset

To train on your own samples: Step 1: Format your dataset to match the structure of our released traces. Step 2: Modify the data path in configs/train.yaml.

---

4. Repository Structure

WordSaladChopper/
├── asset/                  # Documentation assets (PDFs, figures)
├── configs/                # Training configs
├── reproduced/             # Full paper reproduction pipeline
├── sh/                     # Shell entrypoints for generation & training
├── src/                    # Thin wrappers for CLI usage & packaging
├── wscgen/                 # Core Python package
│   ├── chopper.py          # Word Salad detection & chopping logic
│   ├── generate.py         # WSC generator
│   ├── prober.py           # Probe loading / inference helpers
│   ├── pipeline/           # Probe training + evaluation utilities
│   ├── training/           # Training loops for probes
│   └── utils.py            # Shared helper functions
└── pyproject.toml          # Project metadata & dependencies

---

📚 Citation

If you find this work helpful, please cite:

@inproceedings{xie-etal-2025-word,
    title = "Word Salad Chopper: Reasoning Models Waste A Ton Of Decoding Budget On Useless Repetitions, Self-Knowingly",
    author = "Xie, Wenya  and Zhong, Shaochen  and Le, Hoang Anh Duy  and Xu, Zhaozhuo  and Xie, Jianwen  and Liu, Zirui",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing"
}

---

📄 License

Released under the MIT License.

---

Acknowledgment

This project partially builds upon the SkyThought large-scale reasoning framework and draws on the evaluation methodology from Qwen2.5-Math.

• SkyThought: https://github.com/NovaSky-AI/SkyThought
• Qwen2.5-Math: https://github.com/QwenLM/Qwen2.5-Math