Temporal Preference Concepts and their Functions in a Large Language Model
Paper • 2606.05194 • Published
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
layers: list<item: int64>
child 0, item: int64
positions: list<item: string>
child 0, item: string
chunks: list<item: list<item: int64>>
child 0, item: list<item: int64>
child 0, item: int64
d_model: int64
n: int64
layer_convention: string
source_run: string
position_labels: list<item: string>
child 0, item: string
prompts_file: string
n_layers: int64
finished: timestamp[s]
started: timestamp[s]
capture: struct<model_name: string, enable_thinking: bool, max_new_tokens: int64, n_response: int64, batch_si (... 68 chars omitted)
child 0, model_name: string
child 1, enable_thinking: bool
child 2, max_new_tokens: int64
child 3, n_response: int64
child 4, batch_size: int64
child 5, dtype: string
child 6, max_memory_gib: null
child 7, split_layers: int64
n_prompts: int64
model_name: string
to
{'model_name': Value('string'), 'n_layers': Value('int64'), 'd_model': Value('int64'), 'position_labels': List(Value('string')), 'capture': {'model_name': Value('string'), 'enable_thinking': Value('bool'), 'max_new_tokens': Value('int64'), 'n_response': Value('int64'), 'batch_size': Value('int64'), 'dtype': Value('string'), 'max_memory_gib': Value('null'), 'split_layers': Value('int64')}, 'prompts_file': Value('string'), 'n_prompts': Value('int64'), 'started': Value('timestamp[s]'), 'layer_convention': Value('string'), 'finished': Value('timestamp[s]')}
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 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 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
layers: list<item: int64>
child 0, item: int64
positions: list<item: string>
child 0, item: string
chunks: list<item: list<item: int64>>
child 0, item: list<item: int64>
child 0, item: int64
d_model: int64
n: int64
layer_convention: string
source_run: string
position_labels: list<item: string>
child 0, item: string
prompts_file: string
n_layers: int64
finished: timestamp[s]
started: timestamp[s]
capture: struct<model_name: string, enable_thinking: bool, max_new_tokens: int64, n_response: int64, batch_si (... 68 chars omitted)
child 0, model_name: string
child 1, enable_thinking: bool
child 2, max_new_tokens: int64
child 3, n_response: int64
child 4, batch_size: int64
child 5, dtype: string
child 6, max_memory_gib: null
child 7, split_layers: int64
n_prompts: int64
model_name: string
to
{'model_name': Value('string'), 'n_layers': Value('int64'), 'd_model': Value('int64'), 'position_labels': List(Value('string')), 'capture': {'model_name': Value('string'), 'enable_thinking': Value('bool'), 'max_new_tokens': Value('int64'), 'n_response': Value('int64'), 'batch_size': Value('int64'), 'dtype': Value('string'), 'max_memory_gib': Value('null'), 'split_layers': Value('int64')}, 'prompts_file': Value('string'), 'n_prompts': Value('int64'), 'started': Value('timestamp[s]'), 'layer_convention': Value('string'), 'finished': Value('timestamp[s]')}
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.
Planning Temporal Manifolds: Qwen3-14B residual-stream activation subsets
Companion data for Alan's SPAR Fall 2026 project (mentors: Ian Rios-Sialer, Justin Shenk, Shantanu
Darveshi), extending Temporal Concepts and their Shape in Large Language Models (arXiv:2606.05194).
Code and results: snapshot commit afc7430 of CodeReclaimers/SPAR-2026 (shared via the SPAR
repo unrulyabstractions/spar-planning-temporal-manifolds, folder alan/).
Contents
| run | prompts | layers | positions | size |
|---|---|---|---|---|
qwen3-14b_extended_n2000_s0 |
2000 | [22, 37] | T3, R0 | 79 MB |
qwen3-14b_investment_n2000_s0 |
2000 | [22, 37] | T1, T3, T8, R0, R3 | 196 MB |
qwen3-14b_matrix_s0 |
3240 | [22, 26] | T1, T3, R0 | 190 MB |
qwen3.5-27b_extended_n2000_s0 |
2000 | [16, 20, 23, 24, 28, 31, 22, 29, 35, 42, 46, 53, 59] | T0, T1, T2, T3, T4, T5, T6, T7, T8, R0, R1, R2, R3, R4, R5, R6, R7 | 4317 MB |
qwen3.5-27b_families_s0 |
960 | [16, 20, 23, 24, 28, 31, 22, 29, 35, 42, 46, 53, 59] | T0, T1, T2, T3, T4, T5, T6, T7, T8, R0, R1, R2, R3, R4, R5, R6, R7 | 2072 MB |
qwen3.5-27b_families_s0_long |
960 | [16, 20, 23, 24, 28, 31, 22, 29, 35, 42, 46, 53, 59] | T0, T1, T2, T3, T4, T5, T6, T7, T8, R0, R1, R2, R3, R4, R5, R6, R7, M0, E0, MEAN | 2438 MB |
qwen3.5-27b_investment_n2000_s0 |
2000 | [16, 20, 23, 24, 28, 31, 22, 29, 35, 42, 46, 53, 59] | T0, T1, T2, T3, T4, T5, T6, T7, T8, R0, R1, R2, R3, R4, R5, R6, R7 | 4317 MB |
qwen3.5-27b_matrix_s0 |
3240 | [16, 20, 23, 24, 28, 31, 22, 29, 35, 42, 46, 53, 59] | T0, T1, T2, T3, T4, T5, T6, T7, T8, R0, R1, R2, R3, R4, R5, R6, R7 | 6993 MB |
qwen3.5-27b_matrix_s0_long |
3240 | [16, 20, 23, 24, 28, 31, 22, 29, 35, 42, 46, 53, 59] | T0, T1, T2, T3, T4, T5, T6, T7, T8, R0, R1, R2, R3, R4, R5, R6, R7, M0, E0, MEAN | 8228 MB |
qwen3.5-27b_phrasings_n2000_s0 |
2000 | [16, 20, 23, 24, 28, 31, 22, 29, 35, 42, 46, 53, 59] | T0, T1, T2, T3, T4, T5, T6, T7, T8, R0, R1, R2, R3, R4, R5, R6, R7 | 4317 MB |
qwen3.5-27b_units_paired_n2000_s1 |
2000 | [16, 20, 23, 24, 28, 31, 22, 29, 35, 42, 46, 53, 59] | T0, T1, T2, T3, T4, T5, T6, T7, T8, R0, R1, R2, R3, R4, R5, R6, R7 | 4317 MB |
Per run: index.parquet (one row per prompt: prompt text and parameters, generated text, parsed
choice, float32 label logits, p_short), meta.json (model, layer convention, position labels),
subset.json (layers, positions, row chunks), and acts_subset_L{layer}_c{chunk}.safetensors
holding tensor acts of shape [rows, n_positions, 5120] in bfloat16, rows in index.parquet order.
- Model:
Qwen/Qwen3-14B, bfloat16, no-thinking chat template (empty think block pre-filled). - Stored layer
l= residual stream after decoder layerl-1(layer 0 = embeddings); no final norm. - Positions: T0–T8 = the 9-token transition window
<|im_end|>,\n,<|im_start|>,assistant,\n,<think>,\n\n,</think>,\n\n; R0–R3 = the first generated tokens (I,choose,:,a/b). - Activations come from an unpadded per-sample forward pass and are bit-exact against
transformershidden states; a fresh capture reproduces them byte-for-byte.
Loading
from huggingface_hub import snapshot_download
from ptm.subset import load_subset # from the code snapshot; or read the safetensors directly
d = snapshot_download("CodeReclaimers/spar-planning-temporal-manifolds-activations", repo_type="dataset")
X = load_subset(f"{d}/runs/qwen3-14b_investment_n2000_s0", layer=37, position="T3") # float32 [2000, 5120]
Without the package: safetensors.safe_open(file, "pt")["acts"] and subset.json give the same rows.
Full activations (41 layers × 17 positions, 14 GB per run) are not hosted; 1 h 45 on 2 × 16 GB GPUs).scripts/reproduce_all.sh
in the code snapshot regenerates every dataset, capture, and analysis (
- Downloads last month
- 903