Datasets:

anonymousmaharaj
/

laguna-s21-moe-routing

Tags:

License:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

Duplicate

Split (1)

train

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
schema_version: int64
source: string
meta: struct<warning: string, model: string, engine: string, hardware: string, captured_tokens: int64, see (... 30 chars omitted)
  child 0, warning: string
  child 1, model: string
  child 2, engine: string
  child 3, hardware: string
  child 4, captured_tokens: int64
  child 5, seed: null
  child 6, packed: timestamp[s]
dims: struct<layers: int64, experts: int64, top_k: int64>
  child 0, layers: int64
  child 1, experts: int64
  child 2, top_k: int64
sparse_layers: list<item: int64>
  child 0, item: int64
tokens: list<item: struct<i: int64, prompt: int64, domain: string, text: string, experts: list<item: list<it (... 54 chars omitted)
  child 0, item: struct<i: int64, prompt: int64, domain: string, text: string, experts: list<item: list<item: int64>> (... 42 chars omitted)
      child 0, i: int64
      child 1, prompt: int64
      child 2, domain: string
      child 3, text: string
      child 4, experts: list<item: list<item: int64>>
          child 0, item: list<item: int64>
              child 0, item: int64
      child 5, weights: list<item: list<item: double>>
          child 0, item: list<item: double>
              child 0, item: double
metrics: struct<count: list<item: list<item: int64>>, share: list<item: list<item: double>>, mass_share: list (... 296 chars omitted)
  child 0, count: list<item: list<item: int64>>
      child 0, item: list<item: int64>
          child 0, item: int64
  child 1, share: list<item: list<item: double>>
      child 0, item: list<item: double>
          child 0, item: double
  child 2, mass_share: list<item: list<item: double>>
      child 0, item: list<item: double>
          child 0, item: double
  child 3, reap_norm: list<item: list<item: double>>
      child 0, item: list<item: double>
          child 0, item: double
  child 4, domain_share: struct<agentic: list<item: list<item: double>>, code: list<item: list<item: double>>, general: list< (... 110 chars omitted)
      child 0, agentic: list<item: list<item: double>>
          child 0, item: list<item: double>
              child 0, item: double
      child 1, code: list<item: list<item: double>>
          child 0, item: list<item: double>
              child 0, item: double
      child 2, general: list<item: list<item: double>>
          child 0, item: list<item: double>
              child 0, item: double
      child 3, reasoning: list<item: list<item: double>>
          child 0, item: list<item: double>
              child 0, item: double
      child 4, systems: list<item: list<item: double>>
          child 0, item: list<item: double>
              child 0, item: double
domains: list<item: string>
  child 0, item: string
to
{'schema_version': Value('int64'), 'source': Value('string'), 'meta': {'model': Value('string'), 'engine': Value('string'), 'hardware': Value('string'), 'method': Value('string'), 'tokens_profiled': Value('int64'), 'expert_selections': Value('int64'), 'prompts': Value('int64'), 'captured': Value('timestamp[s]'), 'packed': Value('timestamp[s]'), 'profiler_version': Value('string')}, 'dims': {'layers': Value('int64'), 'experts': Value('int64'), 'top_k': Value('int64')}, 'sparse_layers': List(Value('int64')), 'domains': List(Value('string')), 'metrics': {'count': List(List(Value('int64'))), 'share': List(List(Value('float64'))), 'mass_share': List(List(Value('float64'))), 'reap_norm': List(List(Value('float64'))), 'domain_share': {'agentic': List(List(Value('float64'))), 'code': List(List(Value('float64'))), 'general': List(List(Value('float64'))), 'reasoning': List(List(Value('float64'))), 'systems': List(List(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
              schema_version: int64
              source: string
              meta: struct<warning: string, model: string, engine: string, hardware: string, captured_tokens: int64, see (... 30 chars omitted)
                child 0, warning: string
                child 1, model: string
                child 2, engine: string
                child 3, hardware: string
                child 4, captured_tokens: int64
                child 5, seed: null
                child 6, packed: timestamp[s]
              dims: struct<layers: int64, experts: int64, top_k: int64>
                child 0, layers: int64
                child 1, experts: int64
                child 2, top_k: int64
              sparse_layers: list<item: int64>
                child 0, item: int64
              tokens: list<item: struct<i: int64, prompt: int64, domain: string, text: string, experts: list<item: list<it (... 54 chars omitted)
                child 0, item: struct<i: int64, prompt: int64, domain: string, text: string, experts: list<item: list<item: int64>> (... 42 chars omitted)
                    child 0, i: int64
                    child 1, prompt: int64
                    child 2, domain: string
                    child 3, text: string
                    child 4, experts: list<item: list<item: int64>>
                        child 0, item: list<item: int64>
                            child 0, item: int64
                    child 5, weights: list<item: list<item: double>>
                        child 0, item: list<item: double>
                            child 0, item: double
              metrics: struct<count: list<item: list<item: int64>>, share: list<item: list<item: double>>, mass_share: list (... 296 chars omitted)
                child 0, count: list<item: list<item: int64>>
                    child 0, item: list<item: int64>
                        child 0, item: int64
                child 1, share: list<item: list<item: double>>
                    child 0, item: list<item: double>
                        child 0, item: double
                child 2, mass_share: list<item: list<item: double>>
                    child 0, item: list<item: double>
                        child 0, item: double
                child 3, reap_norm: list<item: list<item: double>>
                    child 0, item: list<item: double>
                        child 0, item: double
                child 4, domain_share: struct<agentic: list<item: list<item: double>>, code: list<item: list<item: double>>, general: list< (... 110 chars omitted)
                    child 0, agentic: list<item: list<item: double>>
                        child 0, item: list<item: double>
                            child 0, item: double
                    child 1, code: list<item: list<item: double>>
                        child 0, item: list<item: double>
                            child 0, item: double
                    child 2, general: list<item: list<item: double>>
                        child 0, item: list<item: double>
                            child 0, item: double
                    child 3, reasoning: list<item: list<item: double>>
                        child 0, item: list<item: double>
                            child 0, item: double
                    child 4, systems: list<item: list<item: double>>
                        child 0, item: list<item: double>
                            child 0, item: double
              domains: list<item: string>
                child 0, item: string
              to
              {'schema_version': Value('int64'), 'source': Value('string'), 'meta': {'model': Value('string'), 'engine': Value('string'), 'hardware': Value('string'), 'method': Value('string'), 'tokens_profiled': Value('int64'), 'expert_selections': Value('int64'), 'prompts': Value('int64'), 'captured': Value('timestamp[s]'), 'packed': Value('timestamp[s]'), 'profiler_version': Value('string')}, 'dims': {'layers': Value('int64'), 'experts': Value('int64'), 'top_k': Value('int64')}, 'sparse_layers': List(Value('int64')), 'domains': List(Value('string')), 'metrics': {'count': List(List(Value('int64'))), 'share': List(List(Value('float64'))), 'mass_share': List(List(Value('float64'))), 'reap_norm': List(List(Value('float64'))), 'domain_share': {'agentic': List(List(Value('float64'))), 'code': List(List(Value('float64'))), 'general': List(List(Value('float64'))), 'reasoning': List(List(Value('float64'))), 'systems': List(List(Value('float64')))}}}
              because column names don't match

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.

Laguna S 2.1 — real MoE router profile + per-token route trace

A reproducibility pack: real Mixture-of-Experts router telemetry captured from Laguna S 2.1 (117B MoE) INT4 W4A16 served by vLLM, plus the exact benchmark prompt and harness of the agentic run it was made for.

This dataset is the repro pack for a @superalesha post about an agentic coding run of GLM-5.3 in Claude Code: the model under test built an interactive visualization on top of these files, and everything needed to reproduce the run — data, capture tooling, prompt, launch flags — is published here.

What is in the data

`data/routing.json` — aggregate router profile (real capture)

Model: Laguna S 2.1 (117B MoE) INT4 W4A16
Engine: vLLM, TP=4, enforce-eager, prefill
Hardware: 4x RTX 3090, 96 GB VRAM
47 sparse (MoE) layers x 256 experts, top-10 routing
3,205,231 tokens from 2048 domain-tagged prompts (domains: agentic, code, general, reasoning, systems), 1,506,458,570 (token, expert) selections
Captured 2026-07-23

Method: router selection intercepted in vLLM's FusedMoE sigmoid path via a forward hook (wrapper), and extracted from the workers with collective_rpc; routing is TP-replicated, so worker 0 is authoritative.

Per layer/expert metrics: selection count, share (count normalized within the layer), mass_share (routing-weight share), reap_norm (REAP saliency normalized to the layer max), and domain_share per domain. Uniform routing would give every expert share = 10/256 = 0.0390625; the measured profile does not.

`data/token-trace.json` — per-token route through the stack (real capture)

158 tokens x 47 layers x top-10 expert ids + normalized routing weights, measured on the same Laguna S 2.1 INT4 checkpoint at TP=4 in prefill. The tokens come from three short prompts (Python code, a vLLM serve command line, Russian prose); each token carries its prompt index and domain. The schema also allows a sampled-placeholder variant flagged via source and meta.warning; the shipped file is real-capture with an empty warning.

Full field-by-field schema for both files: data/README.md.

File structure

README.md            this card
PROMPT.md            the published benchmark prompt given to GLM-5.3 (verbatim)
data/
  routing.json       aggregate router profile (real capture)
  token-trace.json   per-token route trace (real capture)
  README.md          schema of both data files
capture/
  prepare-data.py    packs the raw profiler output into routing.json
  make-trace.py      packs the raw trace into token-trace.json
  trace_patch.py     vLLM FusedMoE forward-hook patch that records routes
  capture_trace.py   driver: runs prefill on the rig and dumps the trace
  run-trace-remote.sh  environment + launch wrapper for the capture rig
harness/
  run.sh             Claude Code launch flags for the GLM-5.3 agentic run
  followup.sh        resuming the same session with review instructions

How the data was captured (short version)

vLLM serves the INT4 checkpoint at TP=4 with enforce-eager; a forward hook wrapped around the FusedMoE sigmoid routing path records the top-k expert ids and weights during prefill; collective_rpc pulls the accumulated buffers from the workers (worker 0 is authoritative since routing is TP-replicated). The aggregate profile and the per-token trace come from the same checkpoint and the same capture mechanism. See data/README.md and the scripts in capture/.

License

MIT.

Downloads last month: 24

Total file size:

1.79 MB