sraivante/TinyThinkModel-72M-V1
Text Generation • 72M • Updated • 24
Datasets:
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
created_utc: string
source_directory: string
source_revision_note: string
limits: struct<total_bytes_strictly_less_than: int64, single_file_bytes_strictly_less_than: int64>
child 0, total_bytes_strictly_less_than: int64
child 1, single_file_bytes_strictly_less_than: int64
file_count: int64
total_bytes: int64
total_rows: int64
files: list<item: struct<name: string, bytes: int64, sha256: string, rows: int64, row_groups: int64, source (... 79 chars omitted)
child 0, item: struct<name: string, bytes: int64, sha256: string, rows: int64, row_groups: int64, source: string, u (... 67 chars omitted)
child 0, name: string
child 1, bytes: int64
child 2, sha256: string
child 3, rows: int64
child 4, row_groups: int64
child 5, source: string
child 6, upstream_repo: string
child 7, source_revision: null
child 8, local_mtime_ns: int64
excluded: list<item: null>
child 0, item: null
context: int64
sampling: string
rows_scanned: int64
kinds: struct<long_think: int64, chat: int64, short_cot: int64>
child 0, long_think: int64
child 1, chat: int64
child 2, short_cot: int64
tokenizer_sha256: string
source_manifest_sha256: string
split_rule: string
vocab_size: int64
dtype: string
seed: int64
selected_rows: struct<train: int64, test: int64, validation: int64>
child 0, train: int64
child 1, test: int64
child 2, validation: int64
exact_duplicate_rows_removed: int64
source_tokens: struct<simplescaling__s1K-1.1__data__train-00000-o
...
child 0, item: struct<name: string, tokens: int64, bytes: int64, sha256: string>
child 0, name: string
child 1, tokens: int64
child 2, bytes: int64
child 3, sha256: string
child 1, validation: struct<real_tokens: int64, stored_tokens: int64, padding_tokens: int64, blocks: int64, files: list<i (... 72 chars omitted)
child 0, real_tokens: int64
child 1, stored_tokens: int64
child 2, padding_tokens: int64
child 3, blocks: int64
child 4, files: list<item: struct<name: string, tokens: int64, bytes: int64, sha256: string>>
child 0, item: struct<name: string, tokens: int64, bytes: int64, sha256: string>
child 0, name: string
child 1, tokens: int64
child 2, bytes: int64
child 3, sha256: string
child 2, test: struct<real_tokens: int64, stored_tokens: int64, padding_tokens: int64, blocks: int64, files: list<i (... 72 chars omitted)
child 0, real_tokens: int64
child 1, stored_tokens: int64
child 2, padding_tokens: int64
child 3, blocks: int64
child 4, files: list<item: struct<name: string, tokens: int64, bytes: int64, sha256: string>>
child 0, item: struct<name: string, tokens: int64, bytes: int64, sha256: string>
child 0, name: string
child 1, tokens: int64
child 2, bytes: int64
child 3, sha256: string
packing: string
loss: string
to
{'created_utc': Value('string'), 'seed': Value('int64'), 'dtype': Value('string'), 'context': Value('int64'), 'vocab_size': Value('int64'), 'source_manifest_sha256': Value('string'), 'tokenizer_sha256': Value('string'), 'requested_train_tokens': Value('int64'), 'sampling': Value('string'), 'split_rule': Value('string'), 'packing': Value('string'), 'loss': Value('string'), 'rows_scanned': Value('int64'), 'selected_rows': {'train': Value('int64'), 'test': Value('int64'), 'validation': Value('int64')}, 'exact_duplicate_rows_removed': Value('int64'), 'source_tokens': {'simplescaling__s1K-1.1__data__train-00000-of-00001.parquet/train': Value('int64'), 'open-r1__OpenR1-Math-220k__extended__train-00005-of-00010.parquet/train': Value('int64'), 'open-r1__OpenR1-Math-220k__extended__train-00005-of-00010.parquet/test': Value('int64'), 'open-r1__OpenR1-Math-220k__extended__train-00005-of-00010.parquet/validation': Value('int64'), 'open-thoughts__OpenThoughts-114k__data__train-00003-of-00006.parquet/train': Value('int64'), 'open-thoughts__OpenThoughts-114k__data__train-00003-of-00006.parquet/test': Value('int64'), 'open-thoughts__OpenThoughts-114k__data__train-00003-of-00006.parquet/validation': Value('int64'), 'open-r1__Mixture-of-Thoughts__math__train-00001-of-00004.parquet/train': Value('int64'), 'open-r1__Mixture-of-Thoughts__math__train-00001-of-00004.parquet/validation': Value('int64'), 'open-r1__Mixture-of-Thoughts__math__train-00001-of-00004.parquet/test': Value('int64'), 'open-th
...
int64'), 'open-r1__Mixture-of-Thoughts__science__train-00001-of-00004.parquet/test': Value('int64'), 'a-m-team__AM-DeepSeek-R1-Distilled-1.4M__am_0.9M.parquet/train': Value('int64'), 'a-m-team__AM-DeepSeek-R1-Distilled-1.4M__am_0.9M.parquet/validation': Value('int64'), 'a-m-team__AM-DeepSeek-R1-Distilled-1.4M__am_0.9M.parquet/test': Value('int64'), 'open-r1__Mixture-of-Thoughts__code__train-00002-of-00008.parquet/train': Value('int64'), 'open-r1__Mixture-of-Thoughts__code__train-00002-of-00008.parquet/test': Value('int64'), 'open-r1__Mixture-of-Thoughts__code__train-00002-of-00008.parquet/validation': Value('int64')}, 'kinds': {'long_think': Value('int64'), 'chat': Value('int64'), 'short_cot': Value('int64')}, 'splits': {'train': {'real_tokens': Value('int64'), 'stored_tokens': Value('int64'), 'padding_tokens': Value('int64'), 'blocks': Value('int64'), 'files': List({'name': Value('string'), 'tokens': Value('int64'), 'bytes': Value('int64'), 'sha256': Value('string')})}, 'validation': {'real_tokens': Value('int64'), 'stored_tokens': Value('int64'), 'padding_tokens': Value('int64'), 'blocks': Value('int64'), 'files': List({'name': Value('string'), 'tokens': Value('int64'), 'bytes': Value('int64'), 'sha256': Value('string')})}, 'test': {'real_tokens': Value('int64'), 'stored_tokens': Value('int64'), 'padding_tokens': Value('int64'), 'blocks': Value('int64'), 'files': List({'name': Value('string'), 'tokens': Value('int64'), 'bytes': Value('int64'), 'sha256': Value('string')})}}}
because column names don't match
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 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
created_utc: string
source_directory: string
source_revision_note: string
limits: struct<total_bytes_strictly_less_than: int64, single_file_bytes_strictly_less_than: int64>
child 0, total_bytes_strictly_less_than: int64
child 1, single_file_bytes_strictly_less_than: int64
file_count: int64
total_bytes: int64
total_rows: int64
files: list<item: struct<name: string, bytes: int64, sha256: string, rows: int64, row_groups: int64, source (... 79 chars omitted)
child 0, item: struct<name: string, bytes: int64, sha256: string, rows: int64, row_groups: int64, source: string, u (... 67 chars omitted)
child 0, name: string
child 1, bytes: int64
child 2, sha256: string
child 3, rows: int64
child 4, row_groups: int64
child 5, source: string
child 6, upstream_repo: string
child 7, source_revision: null
child 8, local_mtime_ns: int64
excluded: list<item: null>
child 0, item: null
context: int64
sampling: string
rows_scanned: int64
kinds: struct<long_think: int64, chat: int64, short_cot: int64>
child 0, long_think: int64
child 1, chat: int64
child 2, short_cot: int64
tokenizer_sha256: string
source_manifest_sha256: string
split_rule: string
vocab_size: int64
dtype: string
seed: int64
selected_rows: struct<train: int64, test: int64, validation: int64>
child 0, train: int64
child 1, test: int64
child 2, validation: int64
exact_duplicate_rows_removed: int64
source_tokens: struct<simplescaling__s1K-1.1__data__train-00000-o
...
child 0, item: struct<name: string, tokens: int64, bytes: int64, sha256: string>
child 0, name: string
child 1, tokens: int64
child 2, bytes: int64
child 3, sha256: string
child 1, validation: struct<real_tokens: int64, stored_tokens: int64, padding_tokens: int64, blocks: int64, files: list<i (... 72 chars omitted)
child 0, real_tokens: int64
child 1, stored_tokens: int64
child 2, padding_tokens: int64
child 3, blocks: int64
child 4, files: list<item: struct<name: string, tokens: int64, bytes: int64, sha256: string>>
child 0, item: struct<name: string, tokens: int64, bytes: int64, sha256: string>
child 0, name: string
child 1, tokens: int64
child 2, bytes: int64
child 3, sha256: string
child 2, test: struct<real_tokens: int64, stored_tokens: int64, padding_tokens: int64, blocks: int64, files: list<i (... 72 chars omitted)
child 0, real_tokens: int64
child 1, stored_tokens: int64
child 2, padding_tokens: int64
child 3, blocks: int64
child 4, files: list<item: struct<name: string, tokens: int64, bytes: int64, sha256: string>>
child 0, item: struct<name: string, tokens: int64, bytes: int64, sha256: string>
child 0, name: string
child 1, tokens: int64
child 2, bytes: int64
child 3, sha256: string
packing: string
loss: string
to
{'created_utc': Value('string'), 'seed': Value('int64'), 'dtype': Value('string'), 'context': Value('int64'), 'vocab_size': Value('int64'), 'source_manifest_sha256': Value('string'), 'tokenizer_sha256': Value('string'), 'requested_train_tokens': Value('int64'), 'sampling': Value('string'), 'split_rule': Value('string'), 'packing': Value('string'), 'loss': Value('string'), 'rows_scanned': Value('int64'), 'selected_rows': {'train': Value('int64'), 'test': Value('int64'), 'validation': Value('int64')}, 'exact_duplicate_rows_removed': Value('int64'), 'source_tokens': {'simplescaling__s1K-1.1__data__train-00000-of-00001.parquet/train': Value('int64'), 'open-r1__OpenR1-Math-220k__extended__train-00005-of-00010.parquet/train': Value('int64'), 'open-r1__OpenR1-Math-220k__extended__train-00005-of-00010.parquet/test': Value('int64'), 'open-r1__OpenR1-Math-220k__extended__train-00005-of-00010.parquet/validation': Value('int64'), 'open-thoughts__OpenThoughts-114k__data__train-00003-of-00006.parquet/train': Value('int64'), 'open-thoughts__OpenThoughts-114k__data__train-00003-of-00006.parquet/test': Value('int64'), 'open-thoughts__OpenThoughts-114k__data__train-00003-of-00006.parquet/validation': Value('int64'), 'open-r1__Mixture-of-Thoughts__math__train-00001-of-00004.parquet/train': Value('int64'), 'open-r1__Mixture-of-Thoughts__math__train-00001-of-00004.parquet/validation': Value('int64'), 'open-r1__Mixture-of-Thoughts__math__train-00001-of-00004.parquet/test': Value('int64'), 'open-th
...
int64'), 'open-r1__Mixture-of-Thoughts__science__train-00001-of-00004.parquet/test': Value('int64'), 'a-m-team__AM-DeepSeek-R1-Distilled-1.4M__am_0.9M.parquet/train': Value('int64'), 'a-m-team__AM-DeepSeek-R1-Distilled-1.4M__am_0.9M.parquet/validation': Value('int64'), 'a-m-team__AM-DeepSeek-R1-Distilled-1.4M__am_0.9M.parquet/test': Value('int64'), 'open-r1__Mixture-of-Thoughts__code__train-00002-of-00008.parquet/train': Value('int64'), 'open-r1__Mixture-of-Thoughts__code__train-00002-of-00008.parquet/test': Value('int64'), 'open-r1__Mixture-of-Thoughts__code__train-00002-of-00008.parquet/validation': Value('int64')}, 'kinds': {'long_think': Value('int64'), 'chat': Value('int64'), 'short_cot': Value('int64')}, 'splits': {'train': {'real_tokens': Value('int64'), 'stored_tokens': Value('int64'), 'padding_tokens': Value('int64'), 'blocks': Value('int64'), 'files': List({'name': Value('string'), 'tokens': Value('int64'), 'bytes': Value('int64'), 'sha256': Value('string')})}, 'validation': {'real_tokens': Value('int64'), 'stored_tokens': Value('int64'), 'padding_tokens': Value('int64'), 'blocks': Value('int64'), 'files': List({'name': Value('string'), 'tokens': Value('int64'), 'bytes': Value('int64'), 'sha256': Value('string')})}, 'test': {'real_tokens': Value('int64'), 'stored_tokens': Value('int64'), 'padding_tokens': Value('int64'), 'blocks': Value('int64'), 'files': List({'name': Value('string'), 'tokens': Value('int64'), 'bytes': Value('int64'), 'sha256': Value('string')})}}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
TinyThinkModel-72M-V1-Data
Public, versioned data companion to sraivante/TinyThinkModel-72M-V1. Release public-v1 retains the exact prepared token corpus associated with final step 22,889 and the exact HellaSwag validation snapshot used for evaluation. It does not substitute later source edits for the historical training data. No separate fine-tuning stage was performed.
Snapshot contents and counts
| Split | Selected records | Real tokens | Stored tokens, including tail padding |
|---|---|---|---|
| Train | 670,031 | 1,529,699,784 | 1,529,700,353 |
| Validation | 3,348 | 7,572,390 | 7,572,481 |
| Test | 3,389 | 7,881,773 | 7,882,753 |
The model processed 1,500,053,504 token positions from the prepared corpus. These counts distinguish the prepared training pool from the actual training budget. The binary corpus is 3,090,311,174 bytes before compression; each uncompressed shard is at most 512 MB. The original local pool was 61 Parquet files, 4,188,730 rows, 11,142,811,114 bytes, largest file 2,735,935,999 bytes, within the requested total <20 GB and individual-file <7 GB limits. Only the selected records were prepared. The unselected original rows are not included in this release.
training/*.bin.gz are the original gzip transfer files, verified both by compressed SHA-256 and by streamed decompression against the exact corpus manifest. training/tokenizer/ contains the corpus tokenizer; its BPE JSON is the same as the final model tokenizer. training/split_record_audit.json contains all held-out row positions/hashes plus a bounded train audit, not a complete train-row index. provenance/original_transfer_manifest.json is the original Kaggle transfer inventory; its documentation entries describe that earlier staging folder, whereas this release's complete current inventory is release_files.json.
Schema, preparation and provenance
Original cleaned fields were source (upstream label), kind (record category), prompt, think (possibly empty), and answer. Formatting creates user/assistant message boundaries, optional <think>...</think> text and a document-end token. Output shards are little-endian uint16 token IDs, vocabulary 33,069. Document end separates records; long records continue across 1,024-token windows, and cross-document attention is allowed inside a packed window. Only the final split tail is padded, with pad targets ignored. This is next-token training over both prompt and assistant text.
The split is assigned before tokenizer fitting and packing from SHA-256 of NFKC/casefold/whitespace-normalized prompts: modulo 1000, 0–4 validation, 5–9 test, 10–999 train. Tokenizer fitting uses only training rows and a 128 MB stratified sample. Selected rows come from deterministically shuffled row groups, with per-file train-token quotas proportional to compressed source-file size, seed 20261017. Whole records are retained, so quotas can overshoot. 15,520 exact formatted duplicates were removed. Sampling is not uniform by row; semantic near-duplicates and source errors can remain.
Exact corpus manifest SHA-256: f32e5f3dab44177534cf7457176f83327a02a087c3db2367d771ef64adb2a380. Historical upstream revisions used to construct the local cleaned files were not retained. training/source_manifest.json records their exact local hashes and sizes, with null historical revisions. provenance/upstream_review_2026-10-18.json separately records current upstream revisions and license observations. Neither those later observations nor this repository's upload timestamp is presented as the historical source revision. The source pool is mostly synthetic math/reasoning material and does not provide broad general-language coverage.
Download, restore and inspect
Public download requires no account or token:
from huggingface_hub import snapshot_download
snapshot_download(
"sraivante/TinyThinkModel-72M-V1-Data", repo_type="dataset", revision="public-v1", token=False,
local_dir="tinythink-data", allow_patterns=["training/*", "restore_training_data.py"],
)
python tinythink-data/restore_training_data.py --snapshot tinythink-data --out restored-corpus
The restore script verifies hashes while decompressing and copies the exact tokenizer and manifests. It refuses to overwrite conflicting existing shards. To inspect a restored block:
import numpy as np
from transformers import AutoTokenizer
tokens = np.memmap("restored-corpus/train-0000.bin", dtype="<u2", mode="r")
tok = AutoTokenizer.from_pretrained("restored-corpus/tokenizer", local_files_only=True)
print(tok.decode(tokens[:1024].tolist(), skip_special_tokens=False))
HellaSwag evaluation data
evaluation/hellaswag/validation.parquet is the complete 10,042-example validation split, each with four candidate endings, from Rowan/hellaswag revision 218ec52e09a7e7462a5400043bb9a69a41d06b76. SHA-256: 899813071e1e95efafec90f856e1987d2150fa4d020fc005df6962c259f660cd. Fields: ind, activity_label, ctx_a, ctx_b, ctx, endings, source_id, split, split_type, label. Labels are zero-based choice indices stored as strings.
The original official-harness samples.jsonl records every document, prompt, choice likelihood, label and correctness result. results.json, summary.json, worker source, environment and both GPU-rank tokenization audits accompany it. Evaluation used zero-shot continuation scoring on two T4 GPUs, with no chat template. Raw accuracy is 27.14%; character-length-normalized accuracy is 28.36%. There is no claim of exhaustive training-data decontamination. The pinned HellaSwag README body declares MIT; this source declaration is retained in provenance/public_release_review.json, with attribution and the license text in THIRD_PARTY.md and licenses/MIT.txt.
License and attribution
The owner authorized public visibility on 2026-10-18. The existing exact data files and their hashes are unchanged; public-v1 identifies the updated documentation, while v1 retains the initial-upload history. There was no separate fine-tuning stage.
Copyright (c) 2026 sraivante applies only to original contributions and original selection/arrangement under Apache License 2.0. Third-party records retain their original ownership and component licenses; see LICENSE and THIRD_PARTY.md. The complete collection includes non-commercial/research data and cannot be treated as uniformly Apache-2.0 or commercially unrestricted. Some original per-record license/generator metadata and historical upstream revisions were not retained by the cleaning process. Consult the documented source terms for reuse; publication does not remove those conditions or resolve the missing metadata.
- Downloads last month
- 58
Models trained or fine-tuned on sraivante/TinyThinkModel-72M-V1-Data