data/: the gated 247-query benchmark release using the fake-URL version of the target documents. It includes all 247 IP rewrites and 237 ID rewrites. Ten potentially sensitive ID rewrites are withheld to reduce the risk of harmful redistribution. Researchers may request access to them or generate their own target false claims and ID rewrites.
code/: focused code for the victim search pipeline and quality evaluation.
C-GEO Guard code, weights, and centroids are released separately in the C-GEO Guard repository.

The data intentionally does not include clean/original document text. Each JSONL record contains a query, an IP rewritten target document, release labels, and fake-URL metadata where applicable. Most records also include an ID rewritten target document and its evaluation metadata.

Data files

data/counter_geo_bench_fake_url_247.jsonl: one JSON object per benchmark query, with 247 IP rewrites and 237 released ID rewrites.
data/counter_geo_bench_fake_url_247.sample.json: first three records for quick schema inspection.
data/evaluation_templates/: judge dispatch templates 15 and 16.
DATA_CARD.md: dataset card covering source, fields, intended use, and limitations.

Code Files

The code/ directory preserves the original relative layout for imports:

Victim pipeline: scripts/step2_victim_pipeline.py, plus src/chunker.py, src/retriever.py, src/reranker.py, src/utils.py, and conf/config.yaml.
Evaluation helpers: scripts/phase3_build_quality_eval_tasks.py, scripts/phase3_assemble_quality_eval.py, src/quality.py, and prompt_templates/15_*.md, prompt_templates/16_*.md.

Basic setup

cd code
python -m pip install -e '.[gpu]'

The package can be inspected and syntax-checked on CPU. Running victim synthesis requires GPU access and the models named in conf/config.yaml.

Running the Victim Pipeline

The victim pipeline is the baseline search-and-answer system. It performs chunking, retrieval, reranking, victim-model synthesis, and fact-checking.

The script is:

code/scripts/step2_victim_pipeline.py

It reads a Phase 1 rewrite file from:

code/intermediates/phase1_rewrites.json by default. Data is available in benchmark release. The output is written to:
code/intermediates/phase2_results.json by default

The output filename can also be changed with:

phase2_output=<filename relative to code/intermediates/>

Run the victim pipeline:

cd code
CUDA_VISIBLE_DEVICES=0,1 python scripts/step2_victim_pipeline.py \
  victim.tensor_parallel_size=2

For a custom Phase 1 input/output name:

cd code
CUDA_VISIBLE_DEVICES=0,1 python scripts/step2_victim_pipeline.py \
  phase1_input=phase1_rewrites.json \
  phase2_output=phase2_results.json \
  victim.tensor_parallel_size=2

This baseline requires the victim LLM, embedding model, and reranker named in conf/config.yaml:

Victim LLM: victim.llm_model
Retriever embedding model: embedding.model_name
Reranker: reranker.model_name

The default configuration expects GPUs. In the default setting, victim.tensor_parallel_size=2 uses two GPUs for vLLM and the reranker is placed on the next CUDA device through victim.reranker_device.

C-GEO Guard

The defender implementation, checkpoint, centroids, and usage instructions are maintained in the C-GEO Guard repository. Its evaluation workflow consumes the reranked chunk caches produced by this benchmark harness.

License and responsible use

This repository uses component-specific licenses:

data/ is licensed under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0).
code/ is licensed under the Apache License 2.0.

See LICENSE and LICENSES/ for the applicable terms. Access is gated separately. Users must acknowledge that they are requesting access for defensive research and evaluation. Every record carries an explicit notice identifying it as synthetic benchmark content and stating that its target false claim was generated by a language model. These claims do not represent the authors' advice or views. Released ID rewrites must not be redistributed as an unlabeled web corpus or used to deploy misinformation.

Downloads last month: 31

Total file size:

3.58 MB