ErrOPD-Fashion

Listwise review-ranking data used in ErrOPD: When Does On-Policy Distillation Transfer Teacher Errors? (Xin Li, Hao Jiang, Annan Wang, Yichi Zhang, Chau Yuen). Code: https://github.com/LiXin97/ErrOPD · Project page: https://lixin.ai/ErrOPD · Models and data: ErrOPD collection

Derived from Amazon Reviews 2023 (McAuley Lab), research use only. The review texts and product titles are Amazon customer and catalog content and remain under the terms of the original dataset (https://amazon-reviews-2023.github.io). Our additions (the helpfulness scores, the grouping into 8-review lists, the prompts and the injected errors) are released under CC BY 4.0. See LICENSE.

Each record is a chat (messages: system, user, assistant) plus metadata. The user turn gives a product title and 8 of its reviews, each with a review id; the assistant turn is a strict JSON list of 8 objects with review_id and a 0-10 helpfulness score, in prompt order. Scores were assigned by Gemini 2.5 Pro under a fixed rubric.

Files

File	Config / split	Rows	Use in the paper
train.jsonl	default / train	1,795	student fine-tuning targets and the on-policy distillation prompts
val.jsonl	default / validation	212	held-out prompts for evaluation
teacher_sft/k1_f00_x5.jsonl	teacher_sft / clean	8,975	5 copies of train.jsonl, no injected error (clean teacher control)
teacher_sft/k1_f20_x5.jsonl	teacher_sft / early_close_20pct	8,975	early close (list closed one item early, 7 objects) in 1,828 copies (20.4%); used for the Llama-3.1-8B, Qwen3.5-9B and Olmo-3.1-32B early-close teachers
teacher_sft/k1_f02_x5.jsonl	teacher_sft / early_close_2pct	8,975	early close in 195 copies (2.2%): a student that already makes the error
teacher_sft/dup_f20_x5.jsonl	teacher_sft / duplicated_id_20pct	8,975	duplicated review ID (8 objects, one ID twice) in 1,723 copies (19.2%); used for the Llama-3.1-8B duplicated-ID teacher
teacher_sft/manifest.json			seed (20260930), injection script and sha256 per file

Row counts and injection counts were recounted from the files (an injected copy is a row whose answer differs from the clean answer of its source row). Each product appears in exactly one list; train and validation products are disjoint. Source: the Amazon_Fashion category of Amazon Reviews 2023, built with scripts/data_build_fashion.py; errors injected with scripts/data_inject_fashion.py and scripts/data_inject_fashion_dup.py in the code repository.

Fields

• messages: list of {"role", "content"} turns: system (rubric and output format), user (product title and 8 reviews with ids), assistant (target JSON).
• metadata.item_title: the product title.
• metadata.review_ids: the 8 review ids in prompt order (16-character hashes, not Amazon review or user ids).
• metadata.scores: target 0-10 helpfulness scores in prompt order (Gemini 2.5 Pro).
• metadata.K: list size (8).
• metadata.f2_copy, metadata.f2_row (teacher_sft only): copy index (0-4) and the source row in train.jsonl.

Usage

from datasets import load_dataset

ds = load_dataset("XINLI1997/ErrOPD-Fashion")                    # train (1,795) / validation (212)
teacher = load_dataset("XINLI1997/ErrOPD-Fashion", "teacher_sft") # clean / early_close_20pct / early_close_2pct / duplicated_id_20pct

example = ds["validation"][0]
prompt = [m for m in example["messages"] if m["role"] != "assistant"]  # feed to a chat model
target = example["messages"][-1]["content"]                             # strict JSON list of 8 objects

Personal information

The reviews are public Amazon customer reviews and are reproduced as text. No reviewer names, user ids, timestamps or Amazon review ids are included; review ids are 16-character hashes, and product titles allow re-joining to the original data. Review texts can still mention personal details written by their authors.

Takedown

If you are a rights holder or a reviewer and want content removed, open a discussion on this dataset page or an issue at https://github.com/LiXin97/ErrOPD, and we will remove it.

Citation

@article{li2026erropd,
  title   = {ErrOPD: When Does On-Policy Distillation Transfer Teacher Errors?},
  author  = {Li, Xin and Jiang, Hao and Wang, Annan and Zhang, Yichi and Yuen, Chau},
  journal = {arXiv preprint arXiv:2605.08737},
  year    = {2026}
}

@inproceedings{hou2026bridging,
  title     = {Bridging Language and Items for Retrieval and Recommendation: Benchmarking {LLM}s as Semantic Encoders},
  author    = {Hou, Yupeng and Li, Jiacheng and Fu, Xiangjun and He, Zhankui and Yan, An and Chen, Xiusi and McAuley, Julian},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year      = {2026},
  url       = {https://aclanthology.org/2026.acl-long.147/}
}