File size: 2,736 Bytes
cd5ba1c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
---
license: cc-by-4.0
language:
- en
pretty_name: Redpine Biomedical QA
size_categories:
- n<1K
task_categories:
- question-answering
tags:
- biomedical
- medical
- literature
- retrieval-augmented-generation
- benchmark
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test.jsonl
---

# Redpine Biomedical QA

179 expert-validated biomedical questions, each written from one paper in the Redpine
Science full-text corpus and paired with a gold answer: two to four required claims, a
certainty sentence, the population and setting the claims hold for, and one over-claim
an answer must not make. It is the expert-validated question set of the preprint
[Leveraging a four-quadrant approach for evaluating Redpine Science](https://arxiv.org/abs/2610.07937).

The runner, the judges, the run log and the published verdicts are on GitHub:
[redpine-ai/benchmarks](https://github.com/redpine-ai/benchmarks/tree/main/science/redpine-biomedical-qa).

```python
from datasets import load_dataset

ds = load_dataset("Redpine-AI/redpine-biomedical-qa", split="test")
```

## Fields

| Field | Meaning |
| --- | --- |
| `id` | stable id, `steered/` followed by the source paper's DOI |
| `track` | cardiology (79), neuroimmunology (60), sepsis (20) or immunology (20) |
| `question` | the question as reviewed |
| `required_claims` | the claims a strong answer conveys, in gold order (633 in all) |
| `certainty` | how certain the evidence is |
| `scope` | the population and setting the claims hold for |
| `wrong_if` | the over-claim an answer must not make; null for `steered/10.1177/15230864251401591` |
| `gold` | the full answer key as one string; the four fields above are parsed from it |
| `doi` | the source paper's DOI |
| `access` | `open` (105) or `paywalled` (74), the access class of the source paper |

Every `gold` reads:

    A strong answer conveys: <claim>; <claim>[; <claim>[; <claim>]]. Certainty: <sentence>. Scope: <population and setting>. An answer is wrong if <over-claim>.

## Scoring

The benchmark scores an answer per required claim: a fixed judge model marks each claim
`stated`, `contradicted` or `not_addressed`, and the answer gets claim coverage, a
contradiction rate and a correctness score. The judge prompt and the scoring rules are in
[judging.md](https://github.com/redpine-ai/benchmarks/blob/main/science/redpine-biomedical-qa/judging.md).

## Citation

```bibtex
@misc{dorm2026redpine,
  title         = {Leveraging a four-quadrant approach for evaluating Redpine Science},
  author        = {Dorm, Filip and Vesterbacka, Leonora},
  year          = {2026},
  eprint        = {2610.07937},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2610.07937}
}
```