Datasets:

---

SZLHOLDINGS

/

lean-theorem-tree

Like 0

Follow

SZL Holdings 3

What was broken

The dataset viewer / datasets-server cannot load this dataset. Re-checked on 2026-09-26 against main @ b93a14cd411e7e9168819cf86915f1a02c6bfa2b (last modified 2026-07-22T10:49:34Z):

$ curl -s "https://datasets-server.huggingface.co/is-valid?dataset=SZLHOLDINGS/lean-theorem-tree"
{"preview":false,"viewer":false,"search":false,"filter":false,"statistics":false}

$ curl -s "https://datasets-server.huggingface.co/first-rows?dataset=SZLHOLDINGS/lean-theorem-tree&config=default&split=train"
{"error":"Cannot load the dataset split (in streaming mode) to extract the first rows.","cause_exception":"CastError",
 "cause_message":"Couldn't cast\nmeta: struct<title: string, repo: string, commit: string, commit_date: timestamp[s], ...
 to\n{'canonical_flagship': Value('string'), ... 'schema': Value('string')}\nbecause column names don't match", ...}

The card has no configs:, so the auto-detected default config takes every JSON file in the repo as one train split. The repo has two JSON files with different schemas:

The features were inferred from the receipt, and the data file then fails to cast to them. Even without the receipt, the data file would load as a single row holding a meta struct and a 269-item list. Users expect one row per declaration.

Behavior change for consumers: before 2026-07-22, when SZL_ESTATE_MANAGED.json was added, default/train loaded as 1 row {meta, declarations}. With this change it loads as 269 rows with 9 columns, one per declaration.

What changed

One file changes: README.md. The only edit is a configs: block added to the YAML front matter just before the closing ---. Every other YAML key and the whole markdown body are byte-for-byte unchanged.

@@ -26,6 +26,12 @@
 task_categories:
 - other
 ecosystem-stage: generated-mirror
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/lean_theorem_tree.json
+  field: declarations
 ---

• default / train now reads only data/lean_theorem_tree.json. It uses the JSON builder's field: declarations option, so each of the 269 declarations is one row. The columns are id, name, kind, file, file_sha, line, status, runtime_gate, axiom_deps.
• SZL_ESTATE_MANAGED.json is left out of every config. It is repo-management metadata, not something a dataset user would load. It stays in the repo and can still be downloaded as a raw file.
• The meta block in data/lean_theorem_tree.json is not exposed as a config. With field: meta, the JSON builder returns 3 malformed rows: the nested status_counts dict is split into one scalar per row (247 / 22 / 0), and every other meta value is repeated on each row. Its headline facts are already in the card text: 269 declarations @ c4d13795 from 2026-05-29, Lean v4.14.0-rc1, 247 GREEN / 22 TRACKED, and the thesis and Lean DOIs. Its other values (total_lean_files, the generated timestamp, mathlib_version and the note that line numbers are sampled) are only in the raw file, which is unchanged.
• No features declared. The inferred types are already consistent: id and line are int64, the other scalar columns are string, and axiom_deps is list, which 21 of 269 rows populate. The loaded rows equal the source exactly. That includes the 11 file_sha values that are all digits, which stay strings.

How it was verified

A local snapshot of the repo at b93a14cd was checked by sha256 against the Hub copy. I used datasets 5.0.1 and huggingface_hub 2.0.0 with a fresh, empty datasets cache.

The original README reproduces the server error, with the same cast target: the receipt's 13 keys.

[default streaming=False] ERROR: DatasetGenerationError An error occurred while generating the dataset
[default streaming=False]   caused by: CastError Couldn't cast | meta: struct<title: string, repo: string, commit: string, commit_date: timestamp[s], ...
[default streaming=True] ERROR: CastError Couldn't cast | meta: struct<title: string, ...
CastError ... 'runtime': Value('null'), 'schema': Value('string')}
because column names don't match

New README:

configs: ['default']
[default streaming=False] train num_rows: 269
[default streaming=False] features: {'id': Value('int64'), 'name': Value('string'), 'kind': Value('string'), 'file': Value('string'), 'file_sha': Value('string'), 'line': Value('int64'), 'status': Value('string'), 'runtime_gate': Value('string'), 'axiom_deps': List(Value('string'))}
[default streaming=True] train rows: 269
streaming=False: rows=269 equal_to_source=True
streaming=True: rows=269 equal_to_source=True
builder config: JsonConfig field= declarations data_files= {'train': ['/data/lean_theorem_tree.json']}
remote hf:// streaming field=declarations: rows=269 equal_to_source=True
card configs: [{'config_name': 'default', 'data_files': [{'split': 'train', 'path': 'data/lean_theorem_tree.json'}], 'field': 'declarations'}]
non-configs YAML keys identical: True
markdown body identical: True

equal_to_source=True means the loaded rows equal json.load(...)["declarations"] exactly. The check compares the full list, every row and every field. The count also matches meta.total_declarations (269).

The datasets-server first-rows read path, ds.decode(False) if ds.features else ds followed by islice(..., 101) on the streaming split:

verify_orig split names: ['train']
verify_orig ERROR: CastError because column names don't match
verify_new split names: ['train']
verify_new first-rows-style read: 101 rows; keys: ['id', 'name', 'kind', 'file', 'file_sha', 'line', 'status', 'runtime_gate', 'axiom_deps']

Hub metadata validator (DatasetCard.validate(repo_type="dataset"), a read-only call to /api/validate-yaml):

fresh/README.md validate: OK
staging/README.md validate: OK

What this verification does NOT prove

• It does not prove the Hub datasets-server will build first-rows, parquet, size and statistics for this config. The server runs its own datasets version on Python 3.14. I tested only datasets 5.0.1, locally and over hf://. After merging, re-run /is-valid and /first-rows?config=default&split=train. Those jobs can take a few minutes.
• I did not test older datasets releases. The YAML field pass-through to JsonConfig has existed for a long time, but I did not run it here.
• It says nothing about whether the data is accurate or current. The card itself says this is a snapshot at c4d13795, and the meta.note says line numbers are sampled.
• If a generator regenerates this card from a template without the configs: block, a later regeneration would remove this fix.

🤖 Generated with Claude Code