m38ecaaed/marcap-routing

MarCap Expert-Selection Routing Records

This dataset contains all captured expert-ID sequences from 320 collection segments for each of two MoE models, using requests from allenai/WildChat-1M. Both prefill and decode routing positions are retained for the three captured layers of each model. These are routing records, not conversations, model weights, routing scores, or inference outputs.

Coverage

Model configuration	deepseek	qwen
Model	DeepSeek-V2-Lite-Chat	Qwen3.6-35B-A3B
Captured layers	3, 14, 26	0, 19, 39
Experts / top-k	64 / 6	256 / 8
Segments	320	320
Requests	837383	837383
Prefill routing positions	900951901	758241364
Decode routing positions	395938191	681863610
Total routing positions	1296890092	1440104974

One routing position contains the selections at all three captured layers; it is not counted three times in the totals. The release covers all records in these captured segments, not every layer of the full models or every conversation in the source dataset.

Files

deepseek/segment_000.parquet ... segment_319.parquet
qwen/segment_000.parquet     ... segment_319.parquet
metadata.json

Each Parquet file represents one segment and has one row per request. The records split shown by the Hub contains every released request. No calibration/evaluation labels or experiment-specific partitions are embedded in the data; downstream code can construct request-level splits independently.

Schema

Column	Type	Meaning
request_id	uint32	Stable anonymous request ordinal within one model
prefill_length	uint32	Number of captured prefill routing positions
decode_length	uint32	Number of captured decode routing positions
routes_layer_<layer>	list of fixed-size lists of uint8	Ordered top-k expert IDs at each routing position

There are three routing columns, named by the actual layer numbers above. For each layer, the outer list length equals prefill_length + decode_length. The first prefill_length positions are prefill; the remaining positions are decode. The original top-k slot order is preserved. Expert IDs are model-local expert indices, not destination-rank IDs.

Request IDs range from 0 through 837382 separately for each model. They preserve the request-cache order used for analysis; they are not original dataset row identifiers, conversation identifiers, or arrival timestamps. Do not infer cross-model request correspondence from these IDs alone.

A request with zero captured decode positions is retained with decode_length=0. Captured decode length need not equal the number of tokens in a completion transcript. Request boundaries and token order are retained, but original scheduler interleaving and wall-clock arrival times are not.

Read One Request

Download only the segments needed for an analysis. Use the submission-v1 revision for the submission snapshot. No login or access request is required. For example:

import numpy as np
import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="m38ecaaed/marcap-routing",
    repo_type="dataset",
    filename="deepseek/segment_000.parquet",
    revision="submission-v1",
    token=False,
)
batch = next(pq.ParquetFile(path).iter_batches(batch_size=1))
request = batch.to_pylist()[0]
routes = np.asarray(request["routes_layer_14"], dtype=np.uint8)
prefill = routes[:request["prefill_length"]]
decode = routes[request["prefill_length"]:]

Parquet data use lossless Zstandard compression. The Arrow schema retains unsigned expert IDs and the fixed top-k dimension; readers that expose nested ordinary lists can use the same logical structure. No model execution or GPU is required to read the records.

De-identification and Source

The files contain only anonymous numeric request identifiers, phase lengths, and expert IDs. They do not contain conversation text, token IDs, original request keys, user information, collection timestamps, machine names, or local filesystem paths. The retained routing and length structure is useful for analysis; removal of direct identifiers is not a proof against every possible inference about the underlying inputs.

License and Attribution

This routing database is released under the Open Data Commons Attribution License (ODC-BY) 1.0. It contains information derived from WildChat-1M, which is made available under the same ODC Attribution License. Retain this source attribution and the license notice when redistributing the database.

The records are newly collected expert-selection sequences, not an unchanged copy of WildChat. This database license does not relicense source conversations, model weights or model code. Other applicable rights and upstream terms remain unaffected, including the DeepSeek model license and Qwen model license. No model weights or source conversations are included in this release.

Source dataset reference: Wenting Zhao, Xiang Ren, Jack Hessel, Claire Cardie, Yejin Choi, and Yuntian Deng. WildChat: 1M ChatGPT Interaction Logs in the Wild. ICLR 2024.