README.md

---
license: other
language:
- en
task_categories:
- question-answering
pretty_name: framesLLM
tags:
- rag
- retrieval-augmented-generation
- grounded-generation
- reasoning
- answerability
- evaluation
- chat
size_categories:
- n<1K
dataset_info:
  features:
  - name: id
    dtype: int64
  - name: question
    dtype: string
  - name: gold_answer
    dtype: string
  - name: is_answerable
    dtype: bool
  - name: reasoning_types
    dtype: string
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  splits:
  - name: complete
    num_bytes: 14136218
    num_examples: 824
  - name: balanced
    num_bytes: 3293748
    num_examples: 174
  download_size: 17330257
  dataset_size: 17429966
configs:
- config_name: default
  data_files:
  - split: complete
    path: data/complete-*
  - split: balanced
    path: data/balanced-*
---

# framesLLM

`framesLLM` is a chat-formatted adaptation of the
[FRAMES benchmark](https://huggingface.co/datasets/google/frames-benchmark) for evaluating
retrieval-augmented generation (RAG) systems. It pairs each FRAMES question and reference answer
with a fixed retrieved context embedded in a two-message conversation.

The dataset is designed to test two related capabilities:

- answering multi-hop questions from the supplied documents; and
- recognizing when the supplied context is insufficient to support the reference answer.

The source benchmark contains 824 English questions covering numerical reasoning, tabular
reasoning, multiple constraints, temporal reasoning, and post-processing.

## Dataset origin and transformation

The questions, reference answers, and reasoning labels originate from
[`google/frames-benchmark`](https://huggingface.co/datasets/google/frames-benchmark), introduced in
*Fact, Fetch, and Reason: A Unified Evaluation of Retrieval-Augmented Generation*.

In this adaptation, every example was enriched with retrieved document excerpts and converted to a
chat-oriented representation:

1. a `system` message contains the response instructions and retrieved context, with numbered
   sources; and
2. a `user` message contains the original FRAMES question.

## Splits

| Split | Rows | Answerable | Not answerable | Recommended use |
|---|---:|---:|---:|---|
| `complete` | 824 | 87 | 737 | Full benchmark evaluation |
| `balanced` | 174 | 87 | 87 | Class-balanced answerability evaluation and faster experiments |

`balanced` contains every answerable example from `complete` and an equally sized subset of
non-answerable examples. It is an evaluation subset, not a training split.

## Dataset fields

| Field | Type | Description |
|---|---|---|
| `id` | `int64` | Identifier inherited from the complete adapted dataset. |
| `question` | `string` | Original English multi-hop question. |
| `gold_answer` | `string` | Reference answer from FRAMES. |
| `is_answerable` | `bool` | Whether the embedded retrieved context is sufficient to support the reference answer. |
| `reasoning_types` | `string` | One or more reasoning labels separated by ` | `. |
| `messages` | list of `{role, content}` | Two-message chat representation: one `system` message followed by one `user` message. |

The possible reasoning families are numerical reasoning, tabular reasoning, multiple constraints,
temporal reasoning, and post-processing. A row may combine several families.

## Loading the dataset

```python
from datasets import load_dataset

complete = load_dataset(
    "Mvanypersele/framesLLM",
    split="complete",
)

balanced = load_dataset(
    "Mvanypersele/framesLLM",
    split="balanced",
)
```

To send the prepared conversation to a chat model:

```python
example = complete[0]
messages = example["messages"]
```

The system prompt requires the answer to be grounded exclusively in the supplied context and to end
with a source line. Evaluation code should check the prompt requirements as well as answer
correctness.

## Licensing and attribution

The original FRAMES benchmark is released under the Apache License 2.0. The embedded contexts include
third-party source material, notably Wikipedia-derived excerpts, which may be governed by separate
licenses and attribution requirements. For that reason this repository uses the Hugging Face
metadata value `license: other` rather than applying Apache-2.0 to every stored passage.

Downstream users are responsible for reviewing the licenses of the underlying source documents and
preserving required notices and attribution. The original FRAMES authors and paper must also be
cited.

## Citation

```bibtex
@inproceedings{krishna-etal-2025-fact,
    title = "Fact, Fetch, and Reason: A Unified Evaluation of Retrieval-Augmented Generation",
    author = "Krishna, Satyapriya and Krishna, Kalpesh and Mohananey, Anhad and Schwarcz, Steven and Stambler, Adam and Upadhyay, Shyam and Faruqui, Manaal",
    booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    year = "2025",
    address = "Albuquerque, New Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.naacl-long.243/",
    doi = "10.18653/v1/2025.naacl-long.243",
    pages = "4745--4759",
}
```

## Maintainers

Maintained by [OpenLLM France](https://huggingface.co/OpenLLM-France). For questions about the
adaptation or its construction, open a discussion in this dataset repository.