The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 84, in _split_generators
raise ValueError(
...<2 lines>...
)
ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Hypatia-data
Data released with Hypatia (small open models trained with reinforcement learning and distillation to analyse data in a notebook harness). Code: github.com/fatebreaker/hypatia. Models: huggingface.co/bio-nlp-umass.
Contents
| Folder | What it is |
|---|---|
transcripts/<system>/<benchmark>/run<k>/<problem>.json |
Every graded transcript behind Table 1 of the paper (and GPT-6 Sol at low and medium effort): the agent's notebook (code cells and outputs), its submitted answer and the benchmark grader's verdict. Hypatia rows: the three training runs of the arm shown in each column (arm name in brackets). |
splits/<set>.txt |
Problem IDs of every test set and hold-out set: QRData-Verified (251), MedCalc-Bench-Verified test sample (220), EHRSQL-2024 test sample (300), MedAgentBench (300), BioDSBench (107), Kepler test stars (256); the unused sets used for confirmation (867 EHRSQL-2024 questions, 880 MedCalc-Bench-Verified problems, 157 freshly downloaded Kepler stars) and QRData's 89 questions outside QRData-Verified. |
training_pools/<pool>/train.txt, links.tsv |
The reinforcement-learning training pools and distillation problem sets of the models in Table 1, including Hypatia-9B and Hypatia-4B (one model for all six): the member list of each pool. links.tsv lists files a member reads from shared data (the MIMIC-IV demo database, the MedAgentBench FHIR server's patients, real Kepler light curves), which are not duplicated here; obtain them from their sources with the repository's fetch scripts. |
training_pools/instances/<family>.tar.gz |
Each member instance's problem file (eval_config.json: prompt, answer and grader settings) and its own small data files, one archive per instance family. |
teacher_notebooks/distill27*/ |
The RL-trained Qwen3.6-27B's correct notebooks used for distillation (the paper's 27B teacher; the model itself is not released), with their manifests. |
synthetic_science/<family>.tar.gz |
The synthetic causal-design and synthetic light-curve families used for training and transfer experiments, one archive per family. |
audit/ |
The grounding audit, the notebook-reading (judge) audit and the replication audit; audit/source_data/ holds the source data of every figure and table. |
MANIFEST.json |
File counts and sizes per folder. |
Local file-system paths inside text files are replaced by placeholders (<repo>, <workspace>, scratchpad).
Licences
This dataset combines material derived from several sources; each part keeps its source's licence. Our own additions (transcripts of our models, splits, synthetic-science instances, teacher notebooks, audit tables) are released under CC BY 4.0, except where they reproduce a source's content, in which case that source's terms apply as below.
| Source | Licence | Where it appears |
|---|---|---|
| QRData (Liu et al., 2024) | CC BY-NC 4.0 | QRData transcripts; the qrreal89 pool and split. Non-commercial use only. |
| MedCalc-Bench / MedCalc-Bench-Verified (Khandekar et al., 2024; patient notes from PMC-Patients) | CC BY-SA 4.0 | MedCalc transcripts; the medcalc_ver pool; clinical teacher notebooks. Share-alike. |
| EHRSQL-2024 (Lee et al., 2024) | CC BY 4.0 | EHRSQL transcripts; the ehrsql2024 and clinmix pools; health-record teacher notebooks. |
| MIMIC-IV Clinical Database Demo (PhysioNet) | ODbL 1.0 | Query results inside EHRSQL transcripts. |
| MedAgentBench (Jiang et al., 2025) | MIT | MedAgentBench transcripts; the mablab27mix pools. |
| BioDSBench (Wang et al., 2025) | MIT | BioDSBench transcripts; the bdsgen pool. |
| Kepler mission light curves (MAST) and published catalogues (Kirk et al.; McQuillan et al.; NASA Exoplanet Archive) | public data | Kepler transcripts; the kepreal_mix pool; astronomy teacher notebooks. |
| GPT-6 Sol, gpt-5, gpt-4.1 (OpenAI), Claude Opus 5.5 (Anthropic) outputs | provider terms | Their transcripts, released for evaluation and reproduction only. |
Not included: material derived from GeneBench-Pro case studies, and model weights (released as separate model repositories).
Citation
Please cite the paper and each source benchmark you use.
- Downloads last month
- 80