---
pretty_name: GI-TraceBench
license: apache-2.0
language:
- en
task_categories:
- question-answering
- text-generation
size_categories:
- 10K<n<100K
tags:
- science
- biomedical
- provenance
- evidence-grounded-reasoning
configs:
- config_name: default
  default: true
  data_files:
  - split: train
    path: data/train.parquet
  - split: test
    path: data/test.parquet
---

# GI-TraceBench

GI-TraceBench is a source-linked benchmark for evidence-grounded scientific reasoning
in gastrointestinal oncology. It connects provenance question answering, knowledge
discovery, and hypothesis evaluation through shared Paper, Claim, and Evidence Span
records. The benchmark contains 10,282 English questions across 50 thematic clusters.

## Splits

| Split | Questions | Thematic clusters |
| --- | ---: | ---: |
| train | 8,737 | 42 |
| test | 1,545 | 8 |

The public training split is the finalized adaptation pool used in the paper.
The two splits have disjoint question IDs, normalized question texts, and thematic
clusters. Only `train` and `test` are defined in the default HF configuration.
The benchmark has 5,907 provenance QA, 1,910 knowledge discovery QA, and 2,465
hypothesis evaluation QA instances in total.

## Load the dataset

```python
import json
from datasets import load_dataset

dataset = load_dataset("Sereints/FindLink")
assert len(dataset["train"]) == 8737
assert len(dataset["test"]) == 1545

example = dataset["train"][0]
answer = json.loads(example["answer"])
full_record = json.loads(example["record_json"])
```

The same configuration works with `streaming=True`.

## Files and representations

- `data/train.parquet`, `data/test.parquet`: the default HF loading and viewer
  files. They expose the question, structured answer encoded as a JSON string,
  task metadata, source-ID lists, and the complete record in `record_json`.
- `train.jsonl`, `test.jsonl`: complete structured publication records. Answers
  and nested evidence objects retain their native JSON structure. Read these
  with Python's `json` module; their heterogeneous nested types are not intended
  for automatic Arrow schema inference.
- `source_splits/train.jsonl`, `source_splits/test.jsonl`: the 7,671-question
  source training subset and the same 1,545-question test set. They are download
  artifacts for tracing the source snapshot, excluded from the default HF config.
  The default training pool is the 8,737-question root training file.
- `evidence_store/`: 1,614 paper-ID records, 5,475 Claims, 4,968 Evidence Spans,
  733 associations, and 2,048 hypothesis records. These are structured support
  records, not a collection of complete paper PDFs. They are download artifacts,
  excluded from the default QA loading configuration.
- `dataset_info.json`, `release_alignment.json`, `SHA256SUMS.txt`: representation
  metadata, source checksums, and checksums for every uploaded file respectively.

The default viewer keeps `qa_id`, `qa_group_id`, `qa_family`, `subtask_type`,
`qa_subtype`, `cluster_id`, `language`, `split`, and `question` as strings, and
`claim_ids`, `paper_ids`, `evidence_span_ids`, `association_ids`, and `hypothesis_ids`
as lists of strings. Optional fields may be null. `answer` and `record_json` are
JSON strings. Decoding `record_json` reconstructs the corresponding structured
JSONL record without dropping fields or changing scientific content.

Publication records normalize split and partition metadata to the public
train/test scheme. The frozen source snapshot remains unchanged. This is a
publication view of that snapshot, not a byte-for-byte copy of the source files.
Natural-language uses of terms such as clinical validation are preserved.

## Evaluation and limitations

The dataset supports provenance QA, knowledge discovery, and hypothesis
evaluation under supplied evidence. Reference answers are provided for training
and evaluation; they must not be supplied as model input during held-out testing.
Use `record_json` or the structured JSONL for task-specific assessment fields.

Construction uses AI-assisted processing and structured reference records.
The benchmark is specific to English gastrointestinal oncology literature;
its generated assessment targets and evidence coverage have limitations.
Candidate hypotheses are assessment objects and do not establish clinical
efficacy. This dataset is intended for scientific-reasoning research.

## Licensing and source attribution

The authors' contributions to GI-TraceBench are licensed under the
[Apache License 2.0](LICENSE), as selected for the `Sereints/FindLink` repository.
This grant applies only to rights the authors are entitled to license.
Third-party paper excerpts and source-derived material retain their applicable
original terms and are excluded from any additional grant of rights by this
dataset license. Paper IDs and Evidence Span links support source attribution.
Selecting this license does not establish permission to redistribute source
material; the applicable source terms must also be satisfied.

## Provenance

This publication view is derived from the frozen
`trusted_reasoning_release_20260927` snapshot. `release_alignment.json` records
the source-relative file paths and hashes. No model API credentials or local
authentication files are part of the dataset package.