ShwStone/moonconv-wildchat-prefill

Mooncake-derived WildChat Prefill Workload

A deterministic, ready-to-replay LLM prefill workload. Request arrival shape and target input lengths come from the public Mooncake FAST'25 conversation trace; natural prompt semantics come from prefixes of public WildChat-4.8M conversations. Prompts are materialized as DeepSeek-V3.2 token IDs, so replay does not require redistributing the 3 GB source subset or tokenizing during a benchmark.

This is a derived workload, not an original Mooncake trace release or an AllenAI/DeepSeek release.

What is included

Workload	Requests	Arrival file
formal_0	512	workloads/formal_0_arrivals.jsonl
formal_1	512	workloads/formal_1_arrivals.jsonl
formal_2	512	workloads/formal_2_arrivals.jsonl
screening	128	workloads/screening_arrivals.jsonl
warmup	32	none

The bundle contains 1,696 requests and 17,475,111 input tokens. Each request uses a distinct WildChat conversation. The three formal windows come from three non-overlapping regions of the Mooncake trace.

The main replay records are in workloads/*_requests.jsonl:

{
  "schema_version": 1,
  "request_id": "formal-0-000000",
  "candidate_id": "...",
  "input_length": 63007,
  "output_length": 1,
  "prompt_token_ids": [0, 128803, "..."],
  "prompt_token_ids_sha256": "..."
}

Arrival records contain sequence_index, request_id, base_arrival_offset_ms, the target and actual lengths, and the original Mooncake trace index. Arrival offsets preserve the source timestamp differences exactly and start at zero in each window.

The repository also includes:

• bundle_manifest.json: bundle-level provenance, environment, checks, and file hashes.
• source_manifest.json: exact source revisions and source artifact hashes.
• selection/: deterministic target-to-candidate mapping and selection diagnostics.
• candidate_index/: prompt-free Parquet candidate index for auditing and reselection.
• reports/: machine-readable quality, materialization, benchmark, and reproducibility reports.
• tools/ and tests/: the exact generator and its unit tests.

Frozen sources

• allenai/WildChat-4.8M, revision c827c6df8fcf008219ffaffa4d1dd77491099367; deterministic repository-relative prefix of 22 of 86 shards.
• kvcache-ai/Mooncake, commit e94a0b86ba067455d8b0524eb2cbb5fbac2db024; FAST25-release/traces/conversation_trace.jsonl.
• deepseek-ai/DeepSeek-V3.2, tokenizer revision a7e62ac04ecb2c0a54d736dc46601c5606cf10a6.

The tokenizer repository has no Jinja chat template. Materialization therefore uses the deterministic deepseek_v32 renderer from vLLM 0.26.0, with thinking disabled. The renderer implementation and probe hashes are recorded in source_manifest.json.

Quality

• Bundle status: data_valid.
• Exact input-length matches: 1,459 of 1,696.
• Relative length error: p50 0%, p95 0.008135%, maximum 0.087542%.
• Source subset: 818,576 conversations, 1,828,331 candidate prefixes, and 1,827,737 eligible prefixes.
• Lowest unique-conversation coverage across length bins: 11.93x the required targets.
• Eligible Mooncake trace ratio: 97.888787%.
• Deterministic rebuild reproduced every core artifact byte-for-byte.

All CPU-side source, content, length, arrival, materialization, and reproducibility checks passed. The eight-request replay against the target NPU serving stack has not been run and is explicitly recorded as not_run_requires_npu_service; this release does not claim service-integration validation.

Use

Download the repository:

export HF_HOME=/data/models
HF_HUB_OFFLINE=0 hf download ShwStone/moonconv-wildchat-prefill \
  --repo-type dataset \
  --local-dir /data/models/moonconv-wildchat-prefill

Then read a workload as line-delimited JSON. prompt_token_ids can be sent directly to a compatible serving client, while the corresponding arrival file supplies relative dispatch times.

import json
from pathlib import Path

root = Path("/data/models/moonconv-wildchat-prefill")
with (root / "workloads/formal_0_requests.jsonl").open() as f:
    first_request = json.loads(next(f))

print(first_request["request_id"], first_request["input_length"])

To run the generator tests:

python -m unittest discover -s tests -v

To rebuild, fetch the three frozen upstream revisions, set HF_HOME=/data/models and HF_HUB_OFFLINE=1, then use tools/build_real_prefill_workload.py with build_config.json. The exact source file hashes needed for verification are in source_manifest.json.

Privacy and responsible use

No prompt text, IP-derived identifier, country, state, request header, OpenAI ID, or source timestamp is stored in this repository. Candidate and conversation IDs are one-way hashes. Source coordinates and language labels remain in the audit mapping.

Token IDs are not anonymization: anyone with the matching tokenizer can decode them back into text. WildChat is a public, de-identified dataset, but its conversations may still contain sensitive, offensive, copyrighted, or otherwise undesirable material. Users remain responsible for complying with applicable law, upstream terms, and the rights associated with individual database contents.

License and attribution

The derived database is released under the Open Data Commons Attribution License (ODC-By) 1.0.

Required attribution notice: this database contains transformed information from WildChat-4.8M by AllenAI, revision c827c6df8fcf008219ffaffa4d1dd77491099367, made available under ODC-By 1.0. The transformation selects conversation prefixes, renders them for DeepSeek-V3.2, stores token IDs instead of prompt text, and matches their lengths to Mooncake trace requests.

Mooncake source and trace files at the frozen commit are Apache-2.0; the DeepSeek-V3.2 tokenizer files are MIT. Copies and notices are provided under licenses/ and in THIRD_PARTY_NOTICES.md.

If you use the workload, please also cite the WildChat and Mooncake papers listed in THIRD_PARTY_NOTICES.md.