---
language:
- en
license: apache-2.0
task_categories:
- text-generation
tags:
- cortex
- knowledge-distillation
- code-generation
- seismic-linter
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: example
    path: data.jsonl
---

# Cortex SEIS example

This dataset contains the source documentation and Python evaluation tasks used by the
Cortex consolidation example.
It targets the Python package `seismic-linter==0.2.0` (import name `seismic_linter`).

## Contents

`data.jsonl` contains 11 records:

- One `type="context"` record with the package documentation used to condition the teacher.
- Ten `type="eval"` records with code-generation questions and executable Python tests.

The Hub split is named `example` because the file contains both context and evaluation records.
Each evaluation record has `split="target"`. The file contains no generated training answers,
teacher-distribution caches, model weights, or retention probes.

| Record type | Fields |
| --- | --- |
| `context` | `type`, `id`, `package`, `context` |
| `eval` | `type`, `id`, `split`, `format`, `question`, `tests` |

Evaluation covers static-analysis results and line numbers, file analysis, temporal split
boundaries, and the monotonicity decorator. Questions request a complete `submission.py`;
the `tests` field contains Python `unittest` code that imports that submission.

## Usage

Download the original JSONL without changing its contents:

```python
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="MikeDean/cortex-seis-example",
    repo_type="dataset",
    filename="data.jsonl",
)
print(path)
```

For reproducible runs, also pass `revision="<dataset-commit-hash>"` to `hf_hub_download`.
The Cortex example downloads a pinned version automatically. To override that default,
pass a local path as `--data-file` to both `consolidate.py --stage prepare` and `eval.py`.
The source documentation conditions the teacher;
evaluation questions and tests are kept out of the training-question generation process.

Code scoring requires `seismic-linter==0.2.0` and runs generated Python in an execution
environment supplied by the user. Downloading or reading the dataset does not execute code.

## Provenance

This dataset comes from [Cortex: Consolidating Context into Parametric Memory](https://arxiv.org/abs/2610.20399).

## Citation

[Cortex: Consolidating Context into Parametric Memory](https://arxiv.org/abs/2610.20399).

```bibtex
@misc{zhang2026cortex,
  title = {Cortex: Consolidating Context into Parametric Memory},
  author = {Jintian Zhang and Feng He and Xinnian Liang and Ziwen Xu and Haoming Xu and Haojing Huang and Shuangzhi Wu and Ningyu Zhang},
  year = {2026},
  eprint = {2610.20399},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url = {https://arxiv.org/abs/2610.20399}
}
```