prefill
Datasets
All datasets matching “prefill”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… See the full description on the dataset page: https://huggingface.co/datasets/ShwStone/moonconv-wildchat-prefill.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.… See the full description on the dataset page: https://huggingface.co/datasets/ShwStone/moonconv-wildchat-v4-flash-prefill.rainy-sharegpt-advanced-prefills-filteredprefill-dataset
Prefill Dataset
Long-context tokenized corpus for benchmarking LLM prefill computation with Qwen3-8B. Contains ~10M tokens of copyright-free English text pre-tokenized with character offset mappings for fast position lookup.
Dataset Structure
Files
File
Description
Rows
data/documents.parquet
English documents with token IDs and char offsets
~100-500
data/tasks.parquet
QA, translation, and retrieval tasks
~1K-5K
data/translations.parquet
French… See the full description on the dataset page: https://huggingface.co/datasets/di2ox3/prefill-dataset.c2-sharegpt-advanced-prefills-filteredWM-Prefill-raw-videos
WM-Prefill — raw experiment videos
(中文版: README_zh.md)
The idea in one paragraph. Interactive video world models face a speed/quality dilemma:
a bidirectional video model renders a camera trajectory accurately but is far too slow to
interact with, while a causal (streaming) model is fast but drifts — as it generates
frame after frame it slowly forgets the scene and invents a new one. This project ports the
LLM prefill/decode trick to world models: let the slow bidirectional… See the full description on the dataset page: https://huggingface.co/datasets/Louym/WM-Prefill-raw-videos.