SCRIBES

Code for reproducing the main results of SCRIBES: Web-Scale Script-Based Semi-Structured Data Extraction with Reinforcement Learning. Semi-structured content in HTML tables, lists, and infoboxes accounts for a substantial share of factual data on the web, yet reliably extracting structured information from it remains challenging. Rather than processing each page individually, SCRIBES fine-tunes an LLM with reinforcement learning to write reusable Python extraction scripts, using layout similarity across webpages on the same site as a reward signal so scripts generalize across a site instead of overfitting to one page.

This repository covers two things:

1. Fine-tuning an LLM to generate triple-extraction scripts.
2. Running the fine-tuned model to extract triples from a given HTML page.

Installation

This repo does not include verl itself — clone it separately and apply the patches here:

git clone https://github.com/volcengine/verl.git
cd verl && git checkout d57bfb02b3bf54c4547aaf83fa055f7a1c1cdc6b
git apply /path/to/scribes/verl_patches/constants_ppo.patch
git apply /path/to/scribes/verl_patches/ray_trainer.patch
git apply /path/to/scribes/verl_patches/fsdp_sft_trainer.patch
git apply /path/to/scribes/verl_patches/reward.patch
git apply /path/to/scribes/verl_patches/dp_actor.patch
pip install -e .   # follow verl's own install instructions for your hardware/GPU stack

Then install this repo's own dependencies:

pip install beautifulsoup4 fuzzywuzzy python-Levenshtein munkres scipy vllm requests

Usage

Run these from the root of your verl checkout. Each step's directory has a full README with every flag; this is the minimal path through all three.

1. Build training data (examples/data_preprocess/):

python3 /path/to/scribes/examples/data_preprocess/triple_extraction_group_generalization.py \
    --reuse_test_groups_dir /path/to/an/existing/test_groups_dir \
    --local_dir /path/to/output_dir

Clones SemiBench automatically and builds train.parquet / test.parquet from it.

2. Fine-tune with GRPO (recipe/triple_extraction/):

MODEL_PATH=/path/to/Qwen2.5-14B-Instruct \
TRAIN_FILE=/path/to/output_dir/train.parquet \
VAL_FILE=/path/to/output_dir/test.parquet \
CKPT_DIR=/path/to/checkpoints \
bash /path/to/scribes/recipe/triple_extraction/run_qwen2.5_14b_grpo.sh

3. Extract triples from an HTML page with the fine-tuned model (examples/grpo_trainer/):

python3 /path/to/scribes/examples/grpo_trainer/eval_single_html.py page.html \
    --model_dir /path/to/checkpoints/<checkpoint> \
    --port 8001 --tp 8

Contents

Stage	Directory	What it does
0	verl_patches/	Patches to apply to a stock verl checkout (see examples/data_preprocess/README.md §0 for the pinned commit)
1	examples/data_preprocess/	Builds training data from the public SemiBench benchmark
2	recipe/triple_extraction/	Reward function + GRPO training config/launcher
3	examples/grpo_trainer/	Extracts triples from a single HTML file with the fine-tuned model

Each directory has its own README with exact setup/run instructions.

Citation

@inproceedings{liu2026scribes,
  title={{SCRIBES}: Web-Scale Script-Based Semi-Structured Data Extraction with Reinforcement Learning},
  author={Liu, Shicheng and Sun, Kai and Fu, Lisheng and Chen, Xilun and Zhang, Xinyuan and Lin, Zhaojiang and Shao, Rulin and Liu, Yue and Kumar, Anuj and Yih, Wen-tau and Dong, Xin Luna},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=gQSnEIA3Z3}
}

License

This code is released under CC BY-NC 4.0 (non-commercial use only). It builds training data from SemiBench, which is separately released under the same CC BY-NC 4.0 license — see that repo for details.