Genie: Achieving Human Parity in Content-Grounded Datasets Generation

a three-stage pipeline (content preparation, few-shot LLM generation, filtering) for automatically building content-grounded training datasets

Asaf Yehudai, Boaz Carmeli, Y. Mass, Ofir Arviv, Nathaniel Mills, Assaf Toledo, Eyal Shnarch, Leshem Choshen · arXiv · 2024

In one sentence

Genie builds content-grounded training data automatically by extracting passages, few-shot prompting an LLM to write question-answer pairs or summaries from each passage, and filtering the output with format, NLI-faithfulness and reward-model checks.

Abstract

The lack of high-quality data for content-grounded generation tasks has been identified as a major obstacle to advancing these tasks. To address this gap, we propose Genie, a novel method for automatically generating high-quality content-grounded data. It consists of three stages: (a) Content Preparation, (b) Generation: creating task-specific examples from the content (e.g., question-answer pairs or summaries). (c) Filtering mechanism aiming to ensure the quality and faithfulness of the generated data. We showcase this methodology by generating three large-scale synthetic data, making wishes, for Long-Form Question-Answering (LFQA), summarization, and information extraction. In a human evaluation, our generated data was found to be natural and of high quality. Furthermore, we compare models trained on our data with models trained on human-written data -- ELI5 and ASQA for LFQA and CNN-DailyMail for Summarization. We show that our models are on par with or outperforming models trained on human-generated data and consistently outperforming them in faithfulness. Finally, we applied our method to create LFQA data within the medical domain and compared a model trained on it with models trained on other domains.

Questions this paper answers

can question-answer training data written by a large language model replace a human-written dataset?
how do models finetuned on synthetic content-grounded QA compare with models finetuned on ASQA or ELI5 gold data?
how do I train a long-form QA model when I have documents but no human-written question-answer pairs?
should I pay for human-annotated long-form QA data or generate it with an LLM instead?
Flan-xl trained on 10,000 Genie-generated QA examples reaches ROUGE-L 28.2 and reward 80.3 on ASQA, versus 31.4 and 68.6 for training on human-written ASQA. It also beats human ELI5 training data, which reaches 18.7 ROUGE-L on the ASQA test set.
Holds for: Flan-xl, 10,000 training examples per condition and identical hyperparameters; ASQA-trained model still leads on ROUGE-L and Bert-Score, while synthetic data leads on reward. Falcon-40B is the data generator.
Flan-xl trained on Genie's Wikipedia-from-scratch QA data earns a reward score of 85.1 on ASQA and 26.7 on ELI5. That is above the 72.1 of the ASQA gold answers and just under the 30.3 of the ELI5 gold answers.
Holds for: Reward is the OpenAssistant DeBERTa-v3-large-v2 reference-less model, also used inside Genie's filter, so generator and evaluator share a preference model; Flan-xl base, 10,000 training examples.
Flan-xl trained on Genie's in-domain synthetic QA reaches 81.9 k-Precision and 79.1 ANLI on ASQA, against 67.5 and 55.7 for human ASQA training data and 52.9 and 33.5 for human ELI5 data.
Holds for: Flan-xl, 10,000 examples per condition, ASQA test set. The untrained Flan-xl baseline scores higher still (98.2 k-Precision) with short, near-extractive answers.
can people tell an automatically generated question from one a person actually wrote?
how do LLM-generated questions fare in a forced-choice naturalness comparison against Reddit-sourced and search-log questions?
how do I check whether the questions in my generated dataset read like real user questions?
will reviewers or annotators notice that the questions in my synthetic QA set were machine-written?
In a side-by-side forced-choice test, an expert annotator picked Genie's synthetic question as the human-written one in 72% of ELI5 pairs, 63% of NQ pairs and 49% of ASQA pairs. ELI5 questions come from Reddit posts and NQ questions from Google search logs.
Holds for: 1 non-author expert annotator, 100 question pairs per dataset, each pair grounded in the same content; ASQA questions are crowdsourced.
does throwing away the bad examples from generated training data actually make it better, and by how much?
what do faithfulness and reward-model filters contribute to the human-rated quality of generated content-grounded QA?
how do I raise the quality of an LLM-generated QA dataset without rewriting examples by hand?
is it worth adding a filtering stage to my synthetic data pipeline, or can I train on the raw generations?
Genie's filtering raises human-judged question relevance from 67% to 92% and answers that address the question from 80% to 98%. Answer faithfulness rises from 53% to 76% and overall answer quality from 3.48 to 4.58 on a 5-point Likert scale.
Holds for: 100 randomly sampled content-question-answer triplets from Wikipedia-from-scratch synthetic QA generated by Falcon-40B; the filters are an NLI model and a DeBERTa-v3 reward model at a 0.5 threshold, neither scoring the question directly.
Synthetic QA generated over pre-cleaned NQ passages scores higher on every human quality dimension than synthetic QA generated from freshly crawled Wikipedia. Question clarity is 91% versus 67% and answer faithfulness 88% versus 76%.
Holds for: 100 triplets per condition, single expert annotator, Falcon-40B as generator; both conditions pass the same format, faithfulness and reward filters.
do answers in machine-generated training data stick to the source document better than answers in human-written datasets?
how do k-Precision and ANLI faithfulness scores of models trained on synthetic grounded QA compare with ASQA- and ELI5-trained models?
how do I pick training data that makes a grounded generation model stop making things up?
if hallucination is my main problem, should I train on generated grounded data or on an existing human QA dataset?
Flan-xl trained on Genie's in-domain synthetic QA reaches 81.9 k-Precision and 79.1 ANLI on ASQA, against 67.5 and 55.7 for human ASQA training data and 52.9 and 33.5 for human ELI5 data.
Holds for: Flan-xl, 10,000 examples per condition, ASQA test set. The untrained Flan-xl baseline scores higher still (98.2 k-Precision) with short, near-extractive answers.
On CNN-DailyMail, Flan-xl trained on Genie's Wish-Summarization data scores 28.6 ROUGE-L against 33.3 for training on human CNN-DailyMail summaries, while edging it on reward (97.5 versus 96.5) and k-Precision (98.2 versus 97.0).
Holds for: Flan-xl, CNN-DailyMail test set, 10,000 training examples per condition; parity claim is on Bert-Score, reward and faithfulness metrics, not on ROUGE-L, where human data wins.
does it help to build the generated training data from documents in the same subject area as the task?
how much does domain-matched synthetic grounded QA gain over out-of-domain synthetic data on PubMed-QA?
how do I get a grounded QA model working on medical documents when there is no medical training set?
do I need to generate data from my own domain's documents, or will generated data from Wikipedia passages do?
On PubMed-QA, Flan-xl trained on Genie's medical-domain synthetic QA scores 22.1 ROUGE-L and 81.8 ANLI. Out-of-domain synthetic QA built from NQ passages scores a comparable 22.0 ROUGE-L with higher reward (44.5 versus 39.4) and k-Precision (84.2 versus 78.2).
Holds for: One target domain (medical LFQA on PubMed papers), Flan-xl base, Falcon-40B generator, 10,000 examples per condition; only ANLI favours in-domain data, so the faithfulness comparison is inconclusive.
how much money and time does it take to build a large question-answer dataset with a language model instead of paying people?
what is the per-example API cost of synthesising content-grounded QA relative to expert annotation rates?
how do I budget for creating a few hundred thousand grounded training examples?
can I afford to generate a grounded QA dataset myself rather than commissioning annotators?
Genie generates 10 content-grounded examples in under a minute at $0.02-$0.04 per LLM API call, against roughly $4.45 per expert-written question. At that annotation rate its 300K-example release would have cost over $1M.
Holds for: Cost per question taken from the ASQA paper's reported annotation rate; API costs are those of the models used in 2024 and exclude filtering-model inference and human verification.
does text generated by a language model end up repetitive and narrower in vocabulary than text written by people?
how does the vocd-D lexical diversity of synthesised grounded datasets compare with their human-written counterparts?
how do I tell whether my generated dataset has collapsed into a few repeated phrasings?
should I worry that a 300K-example generated dataset will be less varied than a small human one?
The released Genie datasets span 300K+ samples with vocd-D lexical diversity above their human-written counterparts, for example 76.0 for Wish-Summarization versus 40.1 for CNN-DailyMail. Wish-QA-NQ holds 88K samples against 6,316 for ASQA.
Holds for: vocd-D computed over responses; Wish-QA-NQ's diversity (62.3) is below ASQA's (65.8), while Wish-QA-ELI5 (71.8) exceeds ELI5 (70.9). Sizes reflect what was released, not a cap on what the method can produce.
does the same recipe for generating training data work for summarising documents and pulling facts out of them, not just answering questions?
does a generate-then-filter synthetic data pipeline transfer across content-grounded tasks such as summarization and information extraction?
how do I create training data for document summarisation and information extraction from the same set of documents?
if I already generate synthetic QA this way, can I reuse the pipeline for my summarisation task by swapping the few-shot examples?
On CNN-DailyMail, Flan-xl trained on Genie's Wish-Summarization data scores 28.6 ROUGE-L against 33.3 for training on human CNN-DailyMail summaries, while edging it on reward (97.5 versus 96.5) and k-Precision (98.2 versus 97.0).
Holds for: Flan-xl, CNN-DailyMail test set, 10,000 training examples per condition; parity claim is on Bert-Score, reward and faithfulness metrics, not on ROUGE-L, where human data wins.
Applied to information extraction with only a new set of few-shot examples, Genie's synthetic data lifts Flan-xl from 42.3 to 47.9 ROUGE-L and from 35.9 to 87.4 reward. It beats human Databricks training data on reward and k-Precision but trails it on ROUGE-L, Bert-Score and ANLI.
Holds for: Databricks information-extraction split, 1000 train / 50 test examples, synthetic data generated over Google-NQ passages; the 50-example test set is small.
Genie is a task-agnostic recipe for synthesising content-grounded training data. It extends the generate-then-filter pattern of Self-Instruct and Unnatural Instructions to data that must be grounded in supplied documents rather than in model knowledge.
Holds for: Demonstrated on long-form QA, summarization and information extraction in English, with Falcon-40B and Llama-2-70B as generators, as of the January 2024 preprint; content preparation is rule-based and must be re-implemented per source format.
what work should I read on automatically building training sets for tasks that must stay grounded in supplied documents?
which paper extends instruction-generation pipelines from model-knowledge instructions to content-grounded data synthesis?
where do I start reading about generating training data for retrieval-augmented and document-grounded tasks?
Genie is a task-agnostic recipe for synthesising content-grounded training data. It extends the generate-then-filter pattern of Self-Instruct and Unnatural Instructions to data that must be grounded in supplied documents rather than in model knowledge.
Holds for: Demonstrated on long-form QA, summarization and information extraction in English, with Falcon-40B and Llama-2-70B as generators, as of the January 2024 preprint; content preparation is rule-based and must be re-implemented per source format.
of the two popular long-form question answering datasets built from Reddit and from search queries, which one is better to train on?
which long-form QA training set transfers better across test sets, ASQA or ELI5?
how do I choose between ASQA and ELI5 as the training set for a long-form answer generator?
should I train my long-form answer model on ELI5 or on ASQA?
Genie's experiments provide evidence that ASQA is a higher-quality long-form QA dataset than ELI5. Models trained on ASQA outperform models trained on ELI5 on both test sets, and ELI5 questions were the ones annotators most often mistook synthetic questions for.
Holds for: Comparison is at a matched 10,000 training examples with Flan-xl; ASQA is roughly 6K examples in total against 272K for ELI5, so the quality ordering says nothing about which is more useful at full scale.
does it matter which large language model is used to generate the training data, and does an aligned model beat a plain pretrained one?
how sensitive is content-grounded data synthesis to the choice of generator LLM, aligned versus purely pretrained?
how do I pick which model to use as the generator when synthesising grounded training data?
can I run a synthetic grounded data pipeline with an open pretrained model, or do I need an instruction-aligned one?
Genie is a task-agnostic recipe for synthesising content-grounded training data. It extends the generate-then-filter pattern of Self-Instruct and Unnatural Instructions to data that must be grounded in supplied documents rather than in model knowledge.
Holds for: Demonstrated on long-form QA, summarization and information extraction in English, with Falcon-40B and Llama-2-70B as generators, as of the January 2024 preprint; content preparation is rule-based and must be re-implemented per source format.
Flan-xl trained on 10,000 Genie-generated QA examples reaches ROUGE-L 28.2 and reward 80.3 on ASQA, versus 31.4 and 68.6 for training on human-written ASQA. It also beats human ELI5 training data, which reaches 18.7 ROUGE-L on the ASQA test set.
Holds for: Flan-xl, 10,000 training examples per condition and identical hyperparameters; ASQA-trained model still leads on ROUGE-L and Bert-Score, while synthetic data leads on reward. Falcon-40B is the data generator.

Claims and scope

Common misreadings

Terminology in this paper

Genie
A three-stage method for automatically curating content-grounded datasets: rule-based content preparation, few-shot LLM generation of task examples from each content passage, and filtering by format, NLI-based faithfulness and reward-model quality.
Wish-QA
The family of synthetic question-answering datasets produced by Genie; the bare name also denotes the variant built from scratch by crawling and processing Wikipedia pages (Wikipedia-from-Scratch).
Wish-QA-NQ
Synthetic question-answer data generated over Natural Questions passages, which were originally extracted from Wikipedia by human annotators and are therefore already clean and topically focused.
Wish-QA-MED
Synthetic long-form question-answering data generated over PubMed papers, used as in-domain medical training data.
K-Precision
A lexical faithfulness metric measuring the fraction of words in a generated response that also appear in the grounding content.
content-grounded generation
Tasks where a model must produce a response derived from a supplied document or passage, such as retrieval-augmented question answering, grounded summarization and information extraction.
vocd-D
A lexical diversity measure computed from the vocabulary-size-to-token-count curve, used to compare the vocabulary richness of dataset responses.

How to cite

@inproceedings{DBLP:conf/iclr/YehudaiCMAMSC24,author       = {Asaf Yehudai and
                  Boaz Carmeli and
                  Yosi Mass and
                  Ofir Arviv and
                  Nathaniel Mills and
                  Eyal Shnarch and
                  Leshem Choshen},
  title        = {Achieving Human Parity in Content-Grounded Datasets Generation},
  booktitle    = {The Twelfth International Conference on Learning Representations,
                  {ICLR} 2024, Vienna, Austria, May 7-11, 2024},
  publisher    = {OpenReview.net},
  year         = {2024},
  url          = {https://openreview.net/forum?id=RjYKTQ0L0W},
  timestamp    = {Wed, 07 Aug 2024 01:00:00 +0200},
  biburl       = {https://dblp.org/rec/conf/iclr/YehudaiCMAMSC24.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

References

See the full reference list in the paper.