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
experiment: string
model: string
binary: string
ngl: int64
n_predict: int64
sampling: string
mode: string
n_levels: int64
n_prompts: int64
statistics: struct<spearman_rho: double, spearman_p: double>
child 0, spearman_rho: double
child 1, spearman_p: double
per_prompt: list<item: struct<id: string, level: int64, level_name: string, routing_entropy: double, coalition_s (... 66 chars omitted)
child 0, item: struct<id: string, level: int64, level_name: string, routing_entropy: double, coalition_strength: do (... 54 chars omitted)
child 0, id: string
child 1, level: int64
child 2, level_name: string
child 3, routing_entropy: double
child 4, coalition_strength: double
child 5, diversity_slope: double
child 6, n_prompt_tokens: int64
L1_11: struct<level: string, level_name: string, n_tokens: int64, position_entropy: list<item: double>, las (... 219 chars omitted)
child 0, level: string
child 1, level_name: string
child 2, n_tokens: int64
child 3, position_entropy: list<item: double>
child 0, item: double
child 4, last_token_re: double
child 5, all_token_re: double
child 6, last_token_per_layer: list<item: double>
child 0, item: double
child 7, mean_per_layer: list<item: double>
child 0, item: double
child 8, layer_ids: list<item: int64>
child 0, item: int64
child 9, slope: double
child 10, intercept: double
child 11, r_squared: double
child 12, p_value: double
SL_04: struct<level: strin
...
em: int64
child 9, slope: double
child 10, intercept: double
child 11, r_squared: double
child 12, p_value: double
SR_09: struct<level: string, level_name: string, n_tokens: int64, position_entropy: list<item: double>, las (... 219 chars omitted)
child 0, level: string
child 1, level_name: string
child 2, n_tokens: int64
child 3, position_entropy: list<item: double>
child 0, item: double
child 4, last_token_re: double
child 5, all_token_re: double
child 6, last_token_per_layer: list<item: double>
child 0, item: double
child 7, mean_per_layer: list<item: double>
child 0, item: double
child 8, layer_ids: list<item: int64>
child 0, item: int64
child 9, slope: double
child 10, intercept: double
child 11, r_squared: double
child 12, p_value: double
L5_08: struct<level: string, level_name: string, n_tokens: int64, position_entropy: list<item: double>, las (... 219 chars omitted)
child 0, level: string
child 1, level_name: string
child 2, n_tokens: int64
child 3, position_entropy: list<item: double>
child 0, item: double
child 4, last_token_re: double
child 5, all_token_re: double
child 6, last_token_per_layer: list<item: double>
child 0, item: double
child 7, mean_per_layer: list<item: double>
child 0, item: double
child 8, layer_ids: list<item: int64>
child 0, item: int64
child 9, slope: double
child 10, intercept: double
child 11, r_squared: double
child 12, p_value: double
to
{'L1_11': {'level': Value('string'), 'level_name': Value('string'), 'n_tokens': Value('int64'), 'position_entropy': List(Value('float64')), 'last_token_re': Value('float64'), 'all_token_re': Value('float64'), 'last_token_per_layer': List(Value('float64')), 'mean_per_layer': List(Value('float64')), 'layer_ids': List(Value('int64')), 'slope': Value('float64'), 'intercept': Value('float64'), 'r_squared': Value('float64'), 'p_value': Value('float64')}, 'L3_10': {'level': Value('string'), 'level_name': Value('string'), 'n_tokens': Value('int64'), 'position_entropy': List(Value('float64')), 'last_token_re': Value('float64'), 'all_token_re': Value('float64'), 'last_token_per_layer': List(Value('float64')), 'mean_per_layer': List(Value('float64')), 'layer_ids': List(Value('int64')), 'slope': Value('float64'), 'intercept': Value('float64'), 'r_squared': Value('float64'), 'p_value': Value('float64')}, 'L5_08': {'level': Value('string'), 'level_name': Value('string'), 'n_tokens': Value('int64'), 'position_entropy': List(Value('float64')), 'last_token_re': Value('float64'), 'all_token_re': Value('float64'), 'last_token_per_layer': List(Value('float64')), 'mean_per_layer': List(Value('float64')), 'layer_ids': List(Value('int64')), 'slope': Value('float64'), 'intercept': Value('float64'), 'r_squared': Value('float64'), 'p_value': Value('float64')}, 'SR_09': {'level': Value('string'), 'level_name': Value('string'), 'n_tokens': Value('int64'), 'position_entropy': List(Value('float64')), 'last_token_re': Value('float64'), 'all_token_re': Value('float64'), 'last_token_per_layer': List(Value('float64')), 'mean_per_layer': List(Value('float64')), 'layer_ids': List(Value('int64')), 'slope': Value('float64'), 'intercept': Value('float64'), 'r_squared': Value('float64'), 'p_value': Value('float64')}, 'SL_04': {'level': Value('string'), 'level_name': Value('string'), 'n_tokens': Value('int64'), 'position_entropy': List(Value('float64')), 'last_token_re': Value('float64'), 'all_token_re': Value('float64'), 'last_token_per_layer': List(Value('float64')), 'mean_per_layer': List(Value('float64')), 'layer_ids': List(Value('int64')), 'slope': Value('float64'), 'intercept': Value('float64'), 'r_squared': Value('float64'), 'p_value': Value('float64')}}
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
experiment: string
model: string
binary: string
ngl: int64
n_predict: int64
sampling: string
mode: string
n_levels: int64
n_prompts: int64
statistics: struct<spearman_rho: double, spearman_p: double>
child 0, spearman_rho: double
child 1, spearman_p: double
per_prompt: list<item: struct<id: string, level: int64, level_name: string, routing_entropy: double, coalition_s (... 66 chars omitted)
child 0, item: struct<id: string, level: int64, level_name: string, routing_entropy: double, coalition_strength: do (... 54 chars omitted)
child 0, id: string
child 1, level: int64
child 2, level_name: string
child 3, routing_entropy: double
child 4, coalition_strength: double
child 5, diversity_slope: double
child 6, n_prompt_tokens: int64
L1_11: struct<level: string, level_name: string, n_tokens: int64, position_entropy: list<item: double>, las (... 219 chars omitted)
child 0, level: string
child 1, level_name: string
child 2, n_tokens: int64
child 3, position_entropy: list<item: double>
child 0, item: double
child 4, last_token_re: double
child 5, all_token_re: double
child 6, last_token_per_layer: list<item: double>
child 0, item: double
child 7, mean_per_layer: list<item: double>
child 0, item: double
child 8, layer_ids: list<item: int64>
child 0, item: int64
child 9, slope: double
child 10, intercept: double
child 11, r_squared: double
child 12, p_value: double
SL_04: struct<level: strin
...
em: int64
child 9, slope: double
child 10, intercept: double
child 11, r_squared: double
child 12, p_value: double
SR_09: struct<level: string, level_name: string, n_tokens: int64, position_entropy: list<item: double>, las (... 219 chars omitted)
child 0, level: string
child 1, level_name: string
child 2, n_tokens: int64
child 3, position_entropy: list<item: double>
child 0, item: double
child 4, last_token_re: double
child 5, all_token_re: double
child 6, last_token_per_layer: list<item: double>
child 0, item: double
child 7, mean_per_layer: list<item: double>
child 0, item: double
child 8, layer_ids: list<item: int64>
child 0, item: int64
child 9, slope: double
child 10, intercept: double
child 11, r_squared: double
child 12, p_value: double
L5_08: struct<level: string, level_name: string, n_tokens: int64, position_entropy: list<item: double>, las (... 219 chars omitted)
child 0, level: string
child 1, level_name: string
child 2, n_tokens: int64
child 3, position_entropy: list<item: double>
child 0, item: double
child 4, last_token_re: double
child 5, all_token_re: double
child 6, last_token_per_layer: list<item: double>
child 0, item: double
child 7, mean_per_layer: list<item: double>
child 0, item: double
child 8, layer_ids: list<item: int64>
child 0, item: int64
child 9, slope: double
child 10, intercept: double
child 11, r_squared: double
child 12, p_value: double
to
{'L1_11': {'level': Value('string'), 'level_name': Value('string'), 'n_tokens': Value('int64'), 'position_entropy': List(Value('float64')), 'last_token_re': Value('float64'), 'all_token_re': Value('float64'), 'last_token_per_layer': List(Value('float64')), 'mean_per_layer': List(Value('float64')), 'layer_ids': List(Value('int64')), 'slope': Value('float64'), 'intercept': Value('float64'), 'r_squared': Value('float64'), 'p_value': Value('float64')}, 'L3_10': {'level': Value('string'), 'level_name': Value('string'), 'n_tokens': Value('int64'), 'position_entropy': List(Value('float64')), 'last_token_re': Value('float64'), 'all_token_re': Value('float64'), 'last_token_per_layer': List(Value('float64')), 'mean_per_layer': List(Value('float64')), 'layer_ids': List(Value('int64')), 'slope': Value('float64'), 'intercept': Value('float64'), 'r_squared': Value('float64'), 'p_value': Value('float64')}, 'L5_08': {'level': Value('string'), 'level_name': Value('string'), 'n_tokens': Value('int64'), 'position_entropy': List(Value('float64')), 'last_token_re': Value('float64'), 'all_token_re': Value('float64'), 'last_token_per_layer': List(Value('float64')), 'mean_per_layer': List(Value('float64')), 'layer_ids': List(Value('int64')), 'slope': Value('float64'), 'intercept': Value('float64'), 'r_squared': Value('float64'), 'p_value': Value('float64')}, 'SR_09': {'level': Value('string'), 'level_name': Value('string'), 'n_tokens': Value('int64'), 'position_entropy': List(Value('float64')), 'last_token_re': Value('float64'), 'all_token_re': Value('float64'), 'last_token_per_layer': List(Value('float64')), 'mean_per_layer': List(Value('float64')), 'layer_ids': List(Value('int64')), 'slope': Value('float64'), 'intercept': Value('float64'), 'r_squared': Value('float64'), 'p_value': Value('float64')}, 'SL_04': {'level': Value('string'), 'level_name': Value('string'), 'n_tokens': Value('int64'), 'position_entropy': List(Value('float64')), 'last_token_re': Value('float64'), 'all_token_re': Value('float64'), 'last_token_per_layer': List(Value('float64')), 'mean_per_layer': List(Value('float64')), 'layer_ids': List(Value('int64')), 'slope': Value('float64'), 'intercept': Value('float64'), 'r_squared': Value('float64'), 'p_value': Value('float64')}}
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.
Routing entropy and token position: per-prompt captures, position curves, and the regeneration script
Data and code for the paper on the positional term in mixture-of-experts routing entropy (Shorthill, 2026, Read Routing Entropy at a Fixed Position: A Cross-Model Study of the Prefix Effect in Mixture-of-Experts Routing; Zenodo DOI 10.5281/zenodo.22151499, version 2.0 under concept 10.5281/zenodo.20779602), which revises and extends Sequence Position Explains an Apparent Complexity Gradient in Mixture-of-Experts Routing Entropy (v1.1, Zenodo concept DOI 10.5281/zenodo.20779602). Every number in the paper regenerates from the files here with one script.
What the data show. Per-token routing entropy is depressed over a shared literal prefix and the opening tens of positions and is flat after that; on raw text with no template the depression is confined to the first token. An all-token average therefore dilutes that opening in proportion to prompt length, so it orders a graded prompt suite by length while a fixed-position readout on the same captures removes that order and reverses it at the extremes.
Runs
| Run | Checkpoint | Stack | Input | Files |
|---|---|---|---|---|
| 1 | DeepSeek-V3-0324, Unsloth UD-Q2_K_XL (2-bit) | llama.cpp b8123 fork capture_activations (build f75c4e8), prefill only, ngl 30, ctx 4096 |
Suite L, 168 prompts, chat-template marker strings tokenized as text | results_per_prompt/results_168q_deepseek-v3-0324_prefill.json, capture log |
| 2 | Qwen3.5-397B-A17B, Unsloth UD-Q2_K_XL (2-bit) | same fork and build, ngl 999, flash attention, q8_0 KV, ctx 16384 | Suite L, marker strings as text | results_per_prompt/results_168q_qwen397b_prefill_capture1.json, _capture2.json (determinism repeat) |
| 3 | DeepSeek-R1, Unsloth UD-Q2_K_XL (2-bit) | llama.cpp b8123, prefill only, ngl 30 | Suite L prompts; prompt file not archived | results_per_prompt/results_168q_deepseek-r1_prefill.json |
| 4 | Qwen3.5-35B-A3B-Base, BF16 | Transformers 5.14 qwen3_5_moe, pure-PyTorch DeltaNet, routing hook on every MoE layer |
WikiText-103 train, 1,024-token windows, no template | run4_routing/, run4_curves/ |
Note on names: the Run 1 checkpoint was reported as "DeepSeek V3.1" in the v1.1 paper; the capture
command loaded DeepSeek-V3-0324-UD-Q2_K_XL. Files here use the correct name.
Directory guide
prompts/prompts_168_raw.tsv: Suite L, 168 prompts (id \t text), 12 intended-complexity levels x 14. Level names are recorded per prompt in the results files (level,level_name).prompts/PROMPTS_FROZEN_posconf.tsv: Suite F, the registered 165-prompt factorial grid (5 content categories x 3 length bands x 11), frozen before capture; not yet run.protocol/PROTOCOL_v0.md: the frozen capture protocol for Suite F.results_per_prompt/: one JSON per run. Eachper_promptentry carries the prompt id and level, the rendered token count (n_prompt_tokens), the all-token readout (prefill_re, mean over every rendered prefill position of the layer-averaged normalized router entropy), the final-token readout (last_token_re, absent from the second Qwen capture and from R1), and per-layer values.deepseek-v3-0324_capture_experiment.logis the Run 1 capture log (command, build, devices).position_diagnostic/qwen397b_five_prompt_position_diagnostic.json: full per-position entropy curves for five Suite L prompts (24 to 180 tokens) on Qwen3.5-397B-A17B.run4_routing/: 310.npzfiles, one per WikiText window, withlayers,ids(40 x 1024 x 8 selected expert indices) andgates(their softmax probabilities, not renormalized). 301 windows reach 1,024 tokens; 9 are shorter (952 to 1,016) and are excluded from the curve.fit.logandconvergence.csvcome from the lens-fitting run that produced them.run4_curves/base35_bf16_position_curves_all.npy: the 301 x 1,024 array of top-eight conditional gate entropy (layer-averaged);run4_bootstrap_B2000.npz: prompt-level bootstrap band of the mean curve (B = 2000, seed 20260827);bootstrap_results.json: band widths and the Qwen L1 vs L12 gap CI.regenerate_reported_values.py: recomputes every value reported in the paper (Spearman correlations, level means, Wilcoxon rank-sum tests, bootstrap intervals, diagnostic statistics, Run 4 curve statistics, Suite F composition). Run from this directory:python3 regenerate_reported_values.py(needs numpy and scipy).checksums.txt: SHA-256 for every file in the bundle.
Not included, and why
- Positionwise router tensors for the 168-prompt runs: not retained at capture time; only the per-prompt summaries above survive.
- GGUF, binary, and prompt hashes for Runs 1 to 3, and
nvidia-smioutput: not recorded. - The Run 3 (R1) prompt file: not archived. Its rendered token counts run about ten below Run 1's under the same tokenizer, and its record numbers L9 and L10 in the reverse order of Suite L.
- Run 4 GPU identity: not in the fit log.
Entropy definitions
Runs 1 to 3: normalized Shannon entropy of the softmax over all experts (E = 256 or 512) at each position and layer, averaged over MoE layers. Run 4: top-eight conditional gate entropy, the entropy of the eight stored gate values renormalized to one, divided by log 8. The two are not on a common scale.
License and citation
CC BY 4.0. Cite the paper: Shorthill, J. (2026). Read Routing Entropy at a Fixed Position: A Cross-Model Study of the Prefix Effect in Mixture-of-Experts Routing, version 2.0. Zenodo. doi:10.5281/zenodo.22151499.
- Downloads last month
- 259