Datasets:

lm2445
/

DeRef

license: apache-2.0
pretty_name: DeRef tool-call argument verification sets
language:
  - en
task_categories:
  - text-classification
tags:
  - tool-use
  - function-calling
  - llm-as-a-judge
  - reward-model
size_categories:
  - 1K<n<10K
configs:
  - config_name: fcrb
    data_files:
      - split: test
        path: data/fcrb.jsonl
  - config_name: cfb
    data_files:
      - split: test
        path: data/cfb.jsonl
  - config_name: fintrace
    data_files:
      - split: test
        path: data/fintrace.jsonl
  - config_name: fcrb_relocate
    data_files:
      - split: test
        path: data/fcrb_relocate.jsonl

DeRef tool-call argument verification sets

Evaluation data for the paper "Judges Follow the Reference: The Derivation Gap in Verifying Tool-Call Arguments" . The three benchmark files hold exactly the argument-level analysis sets behind the main-body tables and figures of the paper: tool calls made by an assistant at one step of a task, each paired with a correct or a wrong call, with every argument labelled correct / wrong and assigned a derivation path (where in the context the value can be found). data/fcrb_relocate.jsonl is the controlled set FCRB-Relocate of Figure 2a (see the last section).

from datasets import load_dataset
fcrb = load_dataset("lm2445/DeRef", "fcrb", split="test")

Three configs, one per source benchmark:

config	source	file
`fcrb`	FC-RewardBench (single-turn BFCL v3 tasks)	`data/fcrb.jsonl`
`cfb`	ComplexFuncBench (multi-step booking tasks) with model trajectories from AgentSuite	`data/cfb.jsonl`
`fintrace`	FinTrace (multi-step financial-data tasks)	`data/fintrace.jsonl`
`fcrb_relocate`	controlled set built from FC-RewardBench (Figure 2a)	`data/fcrb_relocate.jsonl`

data/bootstrap_clusters.json lists, per dataset, the full sorted universe of query ids that the paper's cluster bootstrap draws over.

Citation

@misc{ma2026judgesfollowreferencederivation,
      title={Judges Follow the Reference: The Derivation Gap in Verifying Tool-Call Arguments}, 
      author={Linhai Ma and Yan Wang and Lingfei Qian and Yupeng Cao and Victor Gutierrez Basulto},
      year={2026},
      eprint={2610.18497},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.18497}, 
}

Please also cite the source benchmarks (FC-RewardBench, BFCL, ComplexFuncBench, FinTrace) and the AgentSuite trajectories.