Datasets:

malaiwah
/

fruit-phase1-shards

Error code:   StreamingRowsError
Exception:    TypeError
Message:      Couldn't cast array of type
struct<sft_regen: double, sft_magpie: double, replay_fineweb: double, replay_wiki: double>
to
{'glm_regen': Value('float64'), 'glm_magpie': Value('float64'), 'fineweb_edu': Value('float64'), 'wiki_en': Value('float64'), 'wiki_zh': Value('float64'), 'tinystories': Value('float64'), 'reap_calib': Value('float64'), 'spdx': Value('float64'), 'code': Value('float64')}
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, 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 129, 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 489, 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 2818, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, 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 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2303, 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 1852, 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 2149, 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<sft_regen: double, sft_magpie: double, replay_fineweb: double, replay_wiki: double>
              to
              {'glm_regen': Value('float64'), 'glm_magpie': Value('float64'), 'fineweb_edu': Value('float64'), 'wiki_en': Value('float64'), 'wiki_zh': Value('float64'), 'tinystories': Value('float64'), 'reap_calib': Value('float64'), 'spdx': Value('float64'), 'code': Value('float64')}

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.

Fruit Phase-1 tokenized shards

Pre-tokenized inputs for the GLM-5.2-SIQ-Fruit training program. Files are flat NumPy memmaps encoded with the published GLM tokenizer (vocabulary size 154,880), not Arrow/Parquet datasets; the Hugging Face row viewer is therefore not applicable.

The pretraining manifest records 7,546,878,606 tokens across nine source lanes. This public repository contains 7,396,228,297 of those tokens. The 150,650,309-token code lane is intentionally omitted because its gated source is still under redistribution/provenance review.

Pretraining corpus

lane	sampling weight	manifest tokens	published here
GLM-5.2 regen	0.30	4,097,644,506	yes
GLM-5.2 Magpie UltraChat	0.15	634,096,729	yes
FineWeb-Edu	0.20	1,502,660,376	yes
Wikipedia English	0.07	500,343,669	yes
Wikipedia Chinese	0.03	202,624,164	yes
TinyStories	0.08	451,112,884	yes
REAP recall calibration text	0.07	6,717,118	yes
SPDX license text	0.07	1,028,851	yes
code	0.03	150,650,309	no

manifest.json is the machine-readable source of counts and sampling weights. The last 262,144 tokens of each lane are reserved as that lane's fixed validation split and excluded from training sampling.

Apache-2.0 text is deliberately absent from the SPDX lane and was used only as a held-out verbatim-memory needle. The release models' strong MIT continuation and zero Apache overlap are hygiene checks, not general memorization metrics.

SFT corpus

sft/manifest.json defines four weighted memmap lanes:

lane	weight	source-pool tokens	loss mask
`sft_regen`	0.65	210,282,592	assistant-only `.mask.u8`
`sft_magpie`	0.25	90,027,865	assistant-only `.mask.u8`
`replay_fineweb`	0.07	1,502,660,376	full loss
`replay_wiki`	0.03	500,343,669	full loss

Assistant-masked lanes also provide .starts.u64 conversation boundaries. sft/sft-aider.jsonl is the optional Aider-trajectory source used by the trainer's separate trajectory lane.

Contamination notice: any model trained with sft-aider.jsonl is contaminated for Aider/Exercism-style evaluation. Do not report those scores as clean generalization.

Reading the files

import json
from pathlib import Path

import numpy as np

root = Path("fruit-phase1-shards")
manifest = json.loads((root / "manifest.json").read_text())
tokens = np.memmap(root / "tinystories.u32", mode="r", dtype="<u4")
train = tokens[:-manifest["val_tokens"]]
validation = tokens[-manifest["val_tokens"]:]

For masked SFT lanes, read token IDs as <u4, masks as u1, and conversation starts as <u8. The trainer samples without concatenating source files and never crosses the fixed validation tail.

Licensing and redistribution

This is a mixed-source derived dataset and has no single blanket content license. The license: other metadata is intentional. Users must review and comply with each upstream dataset's terms; tokenization does not erase source rights or restrictions. In particular:

the gated code lane is described in manifest.json but not redistributed;
GLM distillation datasets, FineWeb-Edu, Wikipedia, TinyStories, REAP, and SPDX content retain their upstream provenance and terms;
the Aider trajectory file carries the evaluation-contamination warning above.

The preparation code itself is Apache-2.0 and lives in proxy-fruit.

Reproducibility and integrity

fruit_data_prep.py, sft_data_prep.py, and aider_traj_prep.py implement the published formats and validation split. MANIFEST.sha256 authenticates every published data/manifest file except the card, Git attributes, and the integrity manifest itself.

Downloads last month: 48

Total file size:

39.1 GB