Datasets:

caseyallard
/

uor-r4-data

Error code:   StreamingRowsError
Exception:    TypeError
Message:      Couldn't cast array of type
struct<0-31: int64, 128-159: int64, 160-191: int64, 192-223: int64, 224-255: int64, 256-287: int64, 32-63: int64, 64-95: int64, 96-127: int64>
to
{'0-31': Value('int64'), '128-159': Value('int64'), '160-191': Value('int64'), '192-223': Value('int64'), '224-255': Value('int64'), '32-63': Value('int64'), '64-95': Value('int64'), '96-127': Value('int64')}
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2312, in cast_table_to_schema
                  cast_array_to_feature(
                  ~~~~~~~~~~~~~~~~~~~~~^
                      table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                      feature,
                      ^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
                  return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
                                           ~~~~^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2068, in cast_array_to_feature
                  _c(array.field(name) if name in array_fields else null_array, subfeature)
                  ~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
                  return func(array, *args, **kwargs)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2068, in cast_array_to_feature
                  _c(array.field(name) if name in array_fields else null_array, subfeature)
                  ~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
                  return func(array, *args, **kwargs)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2158, in cast_array_to_feature
                  raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
              TypeError: Couldn't cast array of type
              struct<0-31: int64, 128-159: int64, 160-191: int64, 192-223: int64, 224-255: int64, 256-287: int64, 32-63: int64, 64-95: int64, 96-127: int64>
              to
              {'0-31': Value('int64'), '128-159': Value('int64'), '160-191': Value('int64'), '192-223': Value('int64'), '224-255': Value('int64'), '32-63': Value('int64'), '64-95': Value('int64'), '96-127': Value('int64')}

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.

UOR-R4 research data

Project-authored data for the UOR-R4 Geometric Language Model, a pre-alpha Rust research model built on R4/S3/H4 geometry, prime/UOR addressing, exact Z[phi], zeta phases, exact addressed memory and learned typed operators. Models: https://huggingface.co/caseyallard/uor-r4-geometric-214m. Code (MIT): https://github.com/UOR-Foundation/uor-r4. Tracker: https://github.com/UOR-Foundation/uor-r4/issues/2028.

Material authored by the project is published here under MIT, plus one tokenized third-party corpus under its own licence (tinystories/, CDLA-Sharing-1.0). Absolute local paths in the JSON files were replaced by <local>/; the per-directory manifest.json BLAKE3 hashes were computed before that edit, so they match only the binary token/mask files, not the edited JSON.

Start here

Train and score a small native geometric model on a laptop CPU. No GPU and no account are needed.

# 1. Code and the two tools. Install Rust from https://rustup.rs; the repository pins its toolchain.
git clone https://github.com/UOR-Foundation/uor-r4 && cd uor-r4
cargo build --release -p uor-r4-core --bin train-native-prose --bin ablate-prose

# 2. Data: the tokenizer and the tokenized TinyStories split (1.1 GB).
pip install -U "huggingface_hub[cli]"
hf download caseyallard/uor-r4-data --repo-type dataset \
  --include "tinystories/*" --include tokenizer.json --local-dir data

# 3. Train on 2M tokens from token offset 200,000,000. Set --threads to your core count.
mkdir -p run
target/release/train-native-prose --corpus data/tinystories/tinystories_train.u16 \
  --tokenizer data/tokenizer.json --offset 200000000 --tokens 2000000 \
  --lanes 4 --batch-size 256 --seq-len 64 --threads 8 --seed 20260917 \
  --vsa-codes fixed --samples 0 --output run/model.json

# 4. Score the exported artifact (run/model.rgm) on held-out slice A (tokens from 0),
#    with the VSA term and the engram table each ablated.
target/release/ablate-prose --model run/model.rgm \
  --corpus data/tinystories/tinystories_train.u16 --tokenizer data/tokenizer.json \
  --holdout-offset 0 --positions 65536 --vsa-code-mode fixed --ablations vsa,engram

For reference, at main c7fce45b1 on an 8-core Apple laptop, step 3 took 22 s and step 4 took about 7 minutes and reported a baseline of 1.8788 bits per byte over 64,575 positions. The training log ends with an older built-in comparison against a 5-gram model ("Kill Criterion Verdict"); it is not the project's current acceptance test. --samples 0 skips five slow demo completions.

Larger runs of this learner, their arms and their results: vsa-native-test-2026-10-09.
GPU training and evaluation of the 214M geometric stack: the model card and pods-quickstart.md, which runs the project's pod tooling from your own Runpod account.

tokenizer.json: the 4096-entry learned tokenizer used by all 214M checkpoints.
memory-dialogue/: tokenized, project-generated memory dialogues (generator uor_r4_training::milestone_world_v2; copy, associative-recall, relation and other episodes, with response paraphrases written by project sessions).
- mw-balp/, mw-orcp/: train/tokens.u16 (little-endian u16 token ids), train/response_mask.u8 (1 = assistant response token and its end-of-sequence), corpus.json, attempt.json, manifests.
- fine-tune-mix-manifest.json: sanitized manifest of the full 173M-token fine-tune mix (mixB): mw-balp twice, mw-orcp, and chat-v0-p2. Only the two project-generated parts are published; chat-v0-p2 (SmolTalk / UltraChat-derived text) is not republished; obtain it from the sources below.
dev/: tokenized development set (400 project-generated conversations, 42,856 tokens), same format.
panels/: project-authored evaluation panels (conversational-160*, conversational-v2*, everyday-32).
tinystories/: the TinyStories V2 training text tokenized with tokenizer.json (555,385,505 tokens, 1.1 GB, UORT header + u16 ids), the corpus of the native prose learner. Licence: CDLA-Sharing-1.0, inherited from the source; see tinystories/README.md.
MD5SUMS: checksums of the published files.

Not published: held-out panels, third-party-prompt panels, the 8M operator model and compiler artifact, and third-party training text other than tinystories/ (TinyDialogues, SmolLM-Corpus, SmolTalk, UltraChat).