83 lines
2.5 KiB
Markdown
83 lines
2.5 KiB
Markdown
---
|
|
language:
|
|
- zh
|
|
base_model: Qwen/Qwen2.5-3B-Instruct
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- safety
|
|
- toxicity-detection
|
|
- tool-use
|
|
- search-augmented
|
|
---
|
|
|
|
# SeTox-Qwen2.5-3B
|
|
|
|
SeTox-Qwen2.5-3B is a full-SFT model based on `Qwen/Qwen2.5-3B-Instruct` for
|
|
Chinese neologism toxicity detection with optional web-search tool use.
|
|
|
|
## Training
|
|
|
|
- Base model: `Qwen/Qwen2.5-3B-Instruct`
|
|
- Method: full supervised fine-tuning
|
|
- Training data: SeTox tool-use and direct SFT data
|
|
- Epochs: 3
|
|
- Learning rate: `1e-5`
|
|
- Scheduler: linear with warmup ratio `0.05`
|
|
- Context cutoff: `4096`
|
|
|
|
## Evaluation
|
|
|
|
Main SeTox neologism test set, 624 valid samples:
|
|
|
|
| Checkpoint | Accuracy | Unsafe F1 | Unsafe Recall | Safe F1 |
|
|
| --- | ---: | ---: | ---: | ---: |
|
|
| checkpoint-120 | 0.9103 | 0.9402 | 0.8980 | 0.8205 |
|
|
| checkpoint-150 | 0.8974 | 0.9312 | 0.8837 | 0.7987 |
|
|
|
|
Live search and generation settings can introduce small drift during
|
|
reproduction.
|
|
|
|
## Usage
|
|
|
|
See the SeTox code repository for search-augmented inference:
|
|
|
|
```bash
|
|
python scripts/run_inference.py \
|
|
--model_path thu-coai/SeTox-Qwen2.5-3B \
|
|
--search_backend serper \
|
|
--input_path data/eval/neologism_test.json \
|
|
--output_path outputs/neologism_predictions.json
|
|
```
|
|
|
|
For offline smoke tests, use `--mock` or `--search_backend none`.
|
|
|
|
## Limitations
|
|
|
|
This model is intended for research on Chinese neologism toxicity detection. It
|
|
can make mistakes on ambiguous, fast-changing, adversarial, or under-specified
|
|
terms. Search results may contain noisy or unsafe snippets, and downstream users
|
|
should apply appropriate safety review before deployment.
|
|
|
|
## Citation
|
|
|
|
Please cite the SeTox ACL paper:
|
|
|
|
- Paper: https://aclanthology.org/2026.acl-long.1602/
|
|
- DOI: https://doi.org/10.18653/v1/2026.acl-long.1602
|
|
|
|
```bibtex
|
|
@inproceedings{cui-etal-2026-setox,
|
|
title = "{S}e{T}ox: Search-enhanced Reasoning for {LLM}-based Toxicity Detection over {C}hinese Internet Buzzwords",
|
|
author = "Cui, Yiming and Zhang, Qinglin and Su, Xu and Min, Changyu and Hu, Shilin and Huang, Minlie",
|
|
editor = "Che, Wanxiang and Nabende, Joyce and Shutova, Ekaterina and Pilehvar, Mohammad Taher",
|
|
booktitle = "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
|
|
month = jul,
|
|
year = "2026",
|
|
address = "San Diego, California, USA",
|
|
publisher = "Association for Computational Linguistics",
|
|
url = "https://aclanthology.org/2026.acl-long.1602/",
|
|
doi = "10.18653/v1/2026.acl-long.1602",
|
|
pages = "33018--33034"
|
|
}
|
|
```
|