ShwStone/moonconv-wildchat-v4-flash-prefill

Mooncake-derived WildChat DeepSeek-V4-Flash Prefill Workload

A deterministic, ready-to-replay LLM prefill workload for DeepSeek-V4-Flash. 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 V4 token IDs, so a benchmark can submit them directly without decoding, re-applying a chat template, or tokenizing at runtime.

This is a derived workload, not an original release by DeepSeek, AllenAI, or Mooncake. It is the V4 companion to the DeepSeek-V3.2 bundle.

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. Output length is 1 for every request. Every request is an exact natural-prefix length match, and each uses a distinct WildChat conversation; no padding, truncation, text repetition, or copied V3.2 token IDs were used.

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, "...", 128822],
  "prompt_token_ids_sha256": "..."
}

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

The repository also includes provenance manifests, deterministic selection artifacts, a prompt-free Parquet candidate index, machine-readable quality and reproducibility reports, and the exact generator and unit tests.

Frozen sources and encoding

• 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-V4-Flash, revision 60d8d70770c6776ff598c94bb586a859a38244f1.

The frozen V4 tokenizer.json SHA-256 is 8f9f37ca37fdc4f5fd36d5cf4d3b0e8392edb4e894fd10cc0d70b4957c8633cf. The official encoding/encoding_dsv4.py SHA-256 is bdbd57c132a1b3725042323d02b98b9d1df28e5f388f134399555d041f5055e0. Materialization uses the official chat encoding and its <｜Assistant｜></think> generation suffix.

Exact-shape V3.2/V4 contract

The V3.2 and V4 bundles have identical request order, input-length sequence, base-arrival sequence, and Mooncake trace-index sequence:

Invariant	SHA-256
request ID sequence	64ad705ecf25df999b0e77d1d527677f022df703b24688972148105a2c5cccac
input length sequence	9941a7a32d8c94aafb0cb6d2cfa07bf3c27d5299c45475fc3faa33f709f3a816
base arrival sequence	0bc4818432b4b2c05f4f9de44d268dc46ca50fd0a7224f521ca77e0755f5fabf
Mooncake trace-index sequence	f8221efba53a230b2b809bb824873f1f8e5fbcf945ee806e390cd44a5ed70f8b

The V4 prompt-token content SHA-256 is 59d0734bbed76dd547229f6a822571ee9fa591c0aa6bb1e38ec29700afe19d2c; the V3.2 value is 4c4722acf312c7cba296d28a20dddc43f01b626c61977b2b6b911dd0852a715a. All 1,696 per-request prompt hashes differ. See reports/cross_bundle_invariants.json for the machine-readable comparison.

Quality and reproducibility

• Bundle ID: 8ab206ac9f01; status: data_valid.
• Bundle manifest SHA-256: 62af5913ad3c29991b0d2d94737d49b39a3d9fdda31e3030fca3dc88d58d777f.
• Candidate source: 818,576 conversations and 1,828,331 natural prefixes; 1,827,737 prefixes were eligible.
• Exact input-length matches: 1,696 of 1,696.
• All token IDs are integers in [0, 129280); every prompt starts with BOS token 0 and ends with V4 suffix token 128822.
• An independent source retrieval, official encode, selection, and materialization reproduced all 11 core artifacts byte-for-byte.

All CPU-side provenance, content, exact-length, arrival, privacy-field, token-range, and reproducibility checks passed. Replay against the target NPU serving stack has not been run and is recorded as not_run_requires_npu_service; this release does not claim service-integration validation.

Download and replay

export HF_HOME=/data/models
HF_HUB_OFFLINE=0 hf download ShwStone/moonconv-wildchat-v4-flash-prefill \
  --repo-type dataset \
  --local-dir /data/models/moonconv-wildchat-v4-flash-prefill

Read workloads/<group>_requests.jsonl and submit prompt_token_ids directly to the compatible V4 serving client. Do not decode the IDs, add a message template, or tokenize them again. For formal and screening groups, dispatch according to base_arrival_offset_ms in the corresponding arrival file.

import json
from pathlib import Path

root = Path("/data/models/moonconv-wildchat-v4-flash-prefill")
with (root / "workloads/formal_0_requests.jsonl").open() as handle:
    request = json.loads(next(handle))

assert request["input_length"] == len(request["prompt_token_ids"])
# Send request["prompt_token_ids"] directly to the serving client.

The bundle can be validated fully offline:

export HF_HOME=/data/models
export HF_HUB_OFFLINE=1
python tools/build_real_prefill_workload.py validate \
  --config build_config.json \
  --bundle-dir .

Privacy and responsible use

No prompt-text field, IP-derived identifier, country, state, request header, OpenAI ID, or source timestamp is included. Candidate and conversation IDs are one-way hashes; source coordinates and language labels remain for auditing.

Token IDs are not anonymization: anyone with the matching tokenizer can decode them into text. WildChat conversations may contain sensitive, offensive, copyrighted, or otherwise undesirable material. Users remain responsible for applicable law, upstream terms, and 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 natural conversation prefixes, renders them with the frozen official DeepSeek-V4-Flash encoding, stores token IDs rather than prompt text, and exactly matches their lengths to Mooncake trace requests.

Mooncake source and trace files at the frozen commit are Apache-2.0; DeepSeek-V4-Flash tokenizer and encoding files at the frozen revision are MIT. Copies and notices are provided under licenses/ and in THIRD_PARTY_NOTICES.md. Please also cite the WildChat and Mooncake papers listed there.