Home

Awesome

Shepherd: A Critic for Language Model Generation

Tianlu Wang*, Ping Yu*, Xiaoqing Ellen Tan<sup>+</sup>, Sean O'Brien, Ram Pasunuru, Jane Yu, Olga Golovneva, Luke Zettlemoyer, Maryam Fazel-Zarandi, Asli Celikyilmaz

TL;DR: We introduce Shepherd, a language model specifically tuned to critique model responses and suggest refinements, extending beyond the capabilities of an untuned model to identify diverse errors and provide suggestions to remedy them.

<img src="images/overview.png" alt="show" style="zoom:90%;" />

Human annotated feedback

Number of prompts from each dataset

DatasetsNumber of Prompts
Entailment Bank11
Proofwriter162
GSM8K431
PIQA246
CosmosQA143
e-SNLI65
Adversarial NLI68
ECQA118
GPT-3 summarization26
DeFacto29

Error types for human data collection.

Our taxonomy breaks down errors into six specific categories. We require annotators, through our data collection interface, to pinpoint and select these error types accurately, coupled with a well-founded critique. This process allows us to gather data that holds potential for fine-grained training or in-depth evaluation.

Error TypeDescription
ArithmeticError in math calculations.
Coherence and deductionSentences that do not logically follow each other, a summary that lacks a clear topic or conclusion, no structure, steps contradict, etc. This also includes Missing Step that a step in a reasoning/explanation or thought process is missing (typically observed in math or logical reasoning problems).
Consistency with contextInformation about an object (i.e., quantity, characteristics) or a personal named entity does not match information provided in the context/question.
VeracityInformation is not provided in the context and is irrelevant or wrong. For our annotation task rather than needing to look up, please just refer to the correct output which we assume to be the gold answer.
RedundancyExplanation contains redundant information, which even though may be factual, is not required to answer the question and/or repeated in the output.
CommonsenseThe output lacks relations that should be known from the general world. Should be instinctive, without questioning it, based on belief, and accepted by the society, e.g. all ducks are birds.
No errorThe output is correct.

Download data

We inlcude the raw data we collected through Moravia and the data we processed for model training. We also include the data process script we used.

License

The data is under CC-BY-NC 4.0 license.

Citation

Please cite our paper if Shepherd contributes in your work:

@misc{wang2023shepherd,
      title={Shepherd: A Critic for Language Model Generation}, 
      author={Tianlu Wang and Ping Yu and Xiaoqing Ellen Tan and Sean O'Brien and Ramakanth Pasunuru and Jane Dwivedi-Yu and Olga Golovneva and Luke Zettlemoyer and Maryam Fazel-Zarandi and Asli Celikyilmaz},
      year={2023},
      eprint={2308.04592},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}