---
language:
- en
task_categories:
- question-answering
tags:
- biomedical
- knowledge-graph
- reasoning
- benchmark
pretty_name: ImmunoGraphBench2027
configs:
- config_name: medium
  default: true
  data_files:
  - split: test
    path: viewer/medium_final20k_corrected.parquet
- config_name: easy
  data_files:
  - split: test
    path: viewer/easy_final_50000.parquet
- config_name: hard_context_swap
  data_files:
  - split: test
    path: viewer/hard_context_swap_19985.parquet
- config_name: hard_context_quantity
  data_files:
  - split: test
    path: viewer/hard_gold_context_no_advantage_3000.parquet
- config_name: hard_structural_loss
  data_files:
  - split: test
    path: viewer/hard_structural_loss_10000.parquet
---

# ImmunoGraphBench2027

An anonymized review release for studying the contribution of relation-level
context to biomedical knowledge-graph reasoning.

## Included benchmark data

| Configuration | Instances | Description |
|---|---:|---|
| `easy` | 50,000 | Direct factual QA |
| `medium` | 20,000 | Final corrected compositional QA |
| `hard_context_swap` | 19,985 | Relation-context swap intervention |
| `hard_context_quantity` | 3,000 | Gold context has no quantity advantage |
| `hard_structural_loss` | 10,000 | Partial graph-structure removal |

Hard instances are interventions on Medium instances, not an independent pool
of unrelated questions. Do not sum these rows to claim independent examples.
`test` denotes an evaluation collection, not an established train/test split.

Use the configuration selector above to inspect every instance in the Dataset
Viewer. `data/` contains full nested JSONL records (gzip compressed). `viewer/`
contains the same sanitized records in Parquet with stable, readable columns.

## Quick loading

```python
from datasets import load_dataset
import json

ds = load_dataset("ImmunoGraph-Bench/ImmunoGraphBench2027", "medium", split="test")
print(ds[0]["question"])
qa = json.loads(ds[0]["record_json"])
```

The viewer columns are `qa_id`, `question`, `answer`, `level`, `subtype`,
`source_medium_qa_id`, and `record_json`. The last column is the complete QA
object, including its available graph, context and evaluation annotations.
For a non-string answer, the viewer's `answer` is JSON-encoded; the nested
record remains authoritative.

## Record contents and leakage precautions

Medium/Hard records include `question`, `short_answer`, `gold`,
`candidate_subgraph`, and intervention-specific fields where applicable.
Easy records instead use `answer`, `gold_triples`, `gold_path`, and `context`.
Detailed fields can vary by subtype and intervention.

**Do not send the entire QA object to a model.** Gold answers, selected-candidate
annotations and validation metadata are evaluation-only information. Inputs
must be rendered from the fields appropriate to the experimental condition.
Stored structural-loss `model_inputs` are preserved verbatim; other experiments
must use their appropriate condition-specific renderer.

## Minimal evaluation code

The `code/` directory contains dependency-free Python demos:

```sh
python code/evaluate_predictions.py --qa data/Medium/medium_final20k_corrected.jsonl.gz --predictions predictions.jsonl
python -m unittest discover -s code -p 'test_*.py' -v
```

Predictions use one condition per file, with `qa_id` and `answer` (a string or
single-element list). Optional `--aliases aliases.json` accepts an explicitly
provided QA-ID-to-answer-variants mapping. This demo reports simple normalized
exact matching; **it is not the paper's full biomedical alias adjudication**.
`score_frozen.py` separately aggregates existing boolean judgments with an
explicit error policy. See [code documentation](code/README.md).

## Release integrity and anonymization

The published questions, answers, scientific entity IDs, graph relations,
context text and numeric metadata are retained from the frozen source records.
Historical pre-repair questions and internal repair/provenance bookkeeping are
omitted. Absolute `source_file` paths are reduced to basenames. No original Git
history, personal repository links, credentials or internal audit logs are
part of this release. Scientific source citations are retained.

Per-file row counts, sizes and SHA-256 digests are supplied in
`data_manifest.json` and `viewer_manifest.json`.

## Scope and limitations

This initial review deposit contains the five benchmark collections above and
minimal offline evaluation code. It does not yet include the full KG, the full
historical construction pipeline, all model replies, embedding analyses, or
the separate retrieval/random-context/Q2 experiment inputs. It is not a claim
of one-command reproduction of every paper table.

## Source attribution and rights

Records retain available source identifiers and evidence citations. Public
access to this review deposit does not assign a new blanket licence to
third-party database material or literature excerpts. Applicable source terms
remain in effect. A project-wide code/data licence has not yet been assigned;
no unrestricted redistribution rights are asserted here.

This resource is for research evaluation, not clinical decision-making.