CREDIT: evaluation items and training sets

Data for the paper Did the Model Use That? Counterfactual Training Boosts Self-Attribution in Language Models (arXiv: 2610.35295). Code: https://github.com/stanford-oval/CREDIT.

A piece of external information (a user's suggestion, a retrieved passage, a tool's output, ...) is influential if the model gives the answer it supports with the information present but rarely without it. Self-attribution is stating, in the answer, which external information was influential. This dataset holds the evaluation prompts behind the paper's numbers and the relabeled training sets used to train the four CREDIT models.

from datasets import load_dataset

test = load_dataset("CyrusZhouzk/CREDIT", "main", split="test")
train = load_dataset("CyrusZhouzk/CREDIT", "train_qwen3_8b", split="train")

Configs and splits

Config	Splits (rows)
main	dev (1,600), validation (480), validation2 (600), test (1,292), fresh (1,500), large (5,124)
heldout_template_a	validation (400), validation2 (730), test (1,250), fresh (5,500)
heldout_template_b	validation (400), validation2 (730), test (1,250)
heldout_neutral_note	validation (128), validation2 (226), test (410)
competing_two	test (840), fresh (4,000)
competing_three	test (927), fresh (4,383)
frontier_system_notes	test (735)
train_qwen3_8b	train (27,498)
train_qwen3_14b	train (23,108)
train_r1_qwen_7b	train (38,203)
train_r1_llama_8b	train (29,916)

• main: user suggestions on sycophancy-eval questions (TriviaQA, TruthfulQA), plus benchmark passages from FaithEval and RAMDocs in dev, validation, and test. validation and validation2 were used to tune design choices, test was evaluated once, fresh (500 questions) was drawn after the method was frozen, and large (1,708 questions, user suggestions only) adds TruthfulQA questions and was used to compare with frontier models. dev is the development pool used before the splits were drawn. Rows with bench == "pcbench" (PCBench math problems) were built but are not evaluated in the paper.
• heldout_template_a, heldout_template_b: the same questions with external information of six types (user suggestions, tool outputs, another AI's answers, and three types never used in training: question metadata x_meta, a system-prompt note x_sys, claimed expertise x_cred), in two independent templates. heldout_neutral_note: system-prompt notes in neutral wording.
• competing_two, competing_three: two or three pieces of external information per prompt pointing to different answers (pair names the combination; reask_user is the prompt with the scored piece removed).
• frontier_system_notes: the 105 questions used for frontier models with system-prompt notes.
• train_<model>: the relabeled supervised fine-tuning set of each CREDIT model (train_qwen3_8b, train_qwen3_14b, train_r1_qwen_7b for DeepSeek-R1-Distill-Qwen-7B, train_r1_llama_8b for DeepSeek-R1-Distill-Llama-8B).

Fields

Evaluation configs:

Field	Description
bench	source and type of external information: syco (user suggestion), faitheval, ramdocs, pcbench, x_tool (tool output), x_peer (another AI's answer), x_meta, x_sys, x_sysn (neutral note), x_cred; m_*/m3_* for competing pieces
id	question identifier (the question text for sycophancy-eval items)
kind	false: the information points to a wrong answer; true: to the correct answer; plain: no information
mc	multiple-choice question
gold	accepted correct answers (null where not applicable)
n	number of answer options (0 for open-ended questions)
system	system prompt (null if none)
user	user message, including the answer-format instruction; the RELIED ON instruction is appended at evaluation time (see the code)
claim	the external information whose influence is scored (null for plain)

Training configs: prompt (chat-formatted prompt of the base model, including the RELIED ON instruction and the opening <think> tag), completion (the model's own reasoning and answer with the final RELIED ON: line rewritten to match the label), influenced (influence label from counterfactual answers), src (question source and information type: trivia, mmlu, passage, tool, peer), kind (true/false as above).

How the training labels were made

For each training example, the information is influential if the model's answer matches it and at most one of N=4 answers sampled without the information does; it is non-influential if the answer does not match it or at least three counterfactual answers do; other examples are dropped. Answer matching uses an LLM judge (Gemini 3.8 Flash). Only the final RELIED ON: line is rewritten; the reasoning and the answer are the base model's own.

Model outputs, judge decisions, influence labels for the evaluation sets, and human annotations are in the code repository and on OSF.

Models

CREDIT-Qwen3-8B, CREDIT-Qwen3-14B, CREDIT-DeepSeek-R1-Distill-Qwen-7B, CREDIT-DeepSeek-R1-Distill-Llama-8B.

License and terms

The data is derived from sycophancy-eval (Sharma et al.; TriviaQA and TruthfulQA questions), TriviaQA, MMLU, ARC, CommonsenseQA, FaithEval, RAMDocs, and PCBench, which remain under their own licenses, and contains outputs of open models (Qwen3, DeepSeek-R1-Distill) and of a commercial API model used to write passages and alternative answers. It is released for research on model behavior; use is subject to the licenses and terms of the sources (see NOTICE).

Citation

@misc{zhou2026credit,
  title         = {Did the Model Use That? Counterfactual Training Boosts Self-Attribution in Language Models},
  author        = {Zikai Zhou and Jiacheng Sang and Yilin Xu and Yufei Jin and Monica S. Lam},
  year          = {2026},
  eprint        = {2610.35295},
  archivePrefix = {arXiv}
}