# DeepSeek-V4-Flash exact-shape prefill dataset

## Delivery

- Bundle: the repository root after download
- Bundle ID: `8ab206ac9f01`
- Bundle manifest SHA-256: `62af5913ad3c29991b0d2d94737d49b39a3d9fdda31e3030fca3dc88d58d777f`
- Status: `data_valid`
- Requests: 1,696
- Total input tokens: 17,475,111
- Output length: 1 for every request
- Service integration: `not_run_requires_npu_service`

This bundle is a DeepSeek-V4-Flash-specific materialization. It does not copy
V3.2 token IDs. Natural WildChat prefixes were rendered with the frozen
official V4 encoding, tokenized with the V4 tokenizer, and selected only when
their V4 length exactly matched the corresponding V3.2 workload input length.
No padding, truncation, text repetition, runtime decoding, runtime templating,
or runtime tokenization is used.

## Frozen V4 tokenizer and encoding

- Repository: `deepseek-ai/DeepSeek-V4-Flash`
- Revision: `60d8d70770c6776ff598c94bb586a859a38244f1`
- Local frozen directory: `/data/models/deepseek-v4-flash-tokenizer-60d8d707`
- `tokenizer.json` SHA-256: `8f9f37ca37fdc4f5fd36d5cf4d3b0e8392edb4e894fd10cc0d70b4957c8633cf`
- Official `encoding/encoding_dsv4.py` SHA-256: `bdbd57c132a1b3725042323d02b98b9d1df28e5f388f134399555d041f5055e0`
- Tokenizer size including added tokens: 129,280
- Model maximum length: 1,048,576
- Encoding mode: `chat` (`<｜Assistant｜></think>` generation suffix)

The four tests distributed with the official encoding all pass offline. Every
materialized request starts with token ID 0, ends with V4 chat suffix token ID
128822, and stays within `[0, 129280)`.

## Cross-bundle contract

Reference V3.2 bundle: `moonconv_wildchat_v1-fe4f751b6dab`, manifest SHA-256
`adb1926e7a8505c1593de1dafef8f6913b4d9345aaaad6d19370f78ac0914994`.

| Invariant | SHA-256 | V3.2 = V4 |
|---|---|---|
| request ID sequence | `64ad705ecf25df999b0e77d1d527677f022df703b24688972148105a2c5cccac` | yes |
| input length sequence | `9941a7a32d8c94aafb0cb6d2cfa07bf3c27d5299c45475fc3faa33f709f3a816` | yes |
| base arrival sequence | `0bc4818432b4b2c05f4f9de44d268dc46ca50fd0a7224f521ca77e0755f5fabf` | yes |
| Mooncake trace-index sequence | `f8221efba53a230b2b809bb824873f1f8e5fbcf945ee806e390cd44a5ed70f8b` | yes |

All four arrival files and `selection/selected_windows.json` are byte-identical
to V3.2. V4 prompt-token content SHA-256 is
`59d0734bbed76dd547229f6a822571ee9fa591c0aa6bb1e38ec29700afe19d2c`;
the V3.2 value is
`4c4722acf312c7cba296d28a20dddc43f01b626c61977b2b6b911dd0852a715a`.
All 1,696 per-request prompt-token hashes differ between the two bundles,
including requests that happened to select the same source conversation.

The machine-readable comparison is
`reports/cross_bundle_invariants.json`.

## Candidate pool and reproducibility

The deterministic first 22 WildChat shards contain 818,576 source
conversations and 1,828,331 natural user-ending prefixes. Of those, 1,827,737
are within the common 65,536-token limit. Every one of the 1,696 targets found
an exact-length candidate from a distinct conversation.

An independent second candidate match, source retrieval, official V4 encode,
second tokenization, and materialization reproduced 11 core selection and
workload artifacts byte-for-byte. The candidate Parquet index was reused only
after its primary manifest hashes were validated. Details are in
`reports/reproducibility_report.json`.

## Replay

Read `workloads/<group>_requests.jsonl` and submit `prompt_token_ids` directly
to the V4 service. Do not decode, add a template, or tokenize these IDs again.
For formal and screening groups, dispatch according to
`base_arrival_offset_ms` in the matching arrivals file.

```python
import json
from pathlib import Path

bundle = Path(".")
with (bundle / "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.
```

Validate the delivered bundle without network access:

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

The CPU dataset, provenance, exact-length, arrival, privacy-field, token-range,
and reproducibility checks are complete. The target NPU service replay remains
intentionally unclaimed until it is run against the actual checkpoint.