datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
vllm-control-arena
vLLM Main Tasks Dataset
AI coding tasks generated from vLLM git commits
Dataset Description
This dataset contains 6801 coding tasks automatically generated from git commits in the vLLM repository. Each task represents a real-world coding challenge derived from actual development work.
Dataset Structure
The dataset contains the following columns:
commit_hash: The git commit hash
parent_hash: The parent commit hash
commit_title: The original commit… See the full description on the dataset page: https://huggingface.co/datasets/RoganInglis/vllm-control-arena.VLLM_ChartQAvllm-0.28.0-wheels-py312ViLLM-Eval
ViLLM-Eval
We utilize the lm-eval-harness library to conduct evaluations.
This library allows us to efficiently evaluate language models, ensuring robustness and accuracy in our assessments.
Feel free to explore our project and discover the capabilities of the language models we employ.
Install
git clone https://huggingface.co/datasets/vlsp-2023-vllm/ViLLM-Eval
cd ViLLM-Eval
pip install -e .
Basic Usage
# Add trust_remote_code=True if your model is a custom… See the full description on the dataset page: https://huggingface.co/datasets/vlsp-2023-vllm/ViLLM-Eval.MIRB
Benchmarking Multi-Image Understanding in Vision and Language Models: Perception, Knowledge, Reasoning, and Multi-Hop Reasoning
File Structure
├── MIR
|── analogy.json
│── codeu.json
|── dataset_namex.json
└── Images
├── analogy
│ └── image_x.jpg
└──codeu
└── image_x.jpg
JSON Structure
{
"questions": " What is the expected kurtosis of the sequence created by`create_number_sequence(-10, 10)`?\n\n1.… See the full description on the dataset page: https://huggingface.co/datasets/VLLMs/MIRB.details_hosted_vllm____fsx__anton__deepseek-r1-checkpoint_private
Dataset Card for Evaluation run of hosted_vllm//fsx/anton/deepseek-r1-checkpoint
Dataset automatically created during the evaluation run of model hosted_vllm//fsx/anton/deepseek-r1-checkpoint.
The dataset is composed of 15 configuration, each one coresponding to one of the evaluated task.
The dataset has been created from 9 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always… See the full description on the dataset page: https://huggingface.co/datasets/SaylorTwift/details_hosted_vllm____fsx__anton__deepseek-r1-checkpoint_private.MIRB-hfvllm-configs
EXD vLLM Config Profiles
Production configuration profiles for the config-driven serve harness.
Profiles
Config
Model
Notes
qwen2.5-7b
Qwen 2.5 7B Instruct
Baseline 7B
qwen3.6-27b
Qwen 3.6 27B
Mid-size
qwen3.6-35b-a3b
Qwen 3.6 35B A3B
MoE baseline
qwen3.6-35b-a3b-baseline
Qwen 3.6 35B A3B
Baseline sweep
qwen3.6-35b-a3b-mtp0
Qwen 3.6 35B A3B
MTP depth 0
qwen3.6-35b-a3b-mtp3
Qwen 3.6 35B A3B
MTP depth 3
qwen3.6-35b-a3b-throughput
Qwen 3.6 35B… See the full description on the dataset page: https://huggingface.co/datasets/EXDai/vllm-configs.vllm-configs
EXD vLLM Config Profiles
Production configuration profiles for the config-driven serve harness at projects/serve/.
Usage
On atom (GPU machine):
cd ~/EXD
./scripts/up.sh <config-name>
Profiles
Config
Model
Notes
qwen2.5-7b
Qwen 2.5 7B Instruct
Baseline 7B
qwen3.6-27b
Qwen 3.6 27B
Mid-size
qwen3.6-35b-a3b
Qwen 3.6 35B A3B
MoE baseline
qwen3.6-35b-a3b-baseline
Qwen 3.6 35B A3B
Baseline sweep
qwen3.6-35b-a3b-mtp0
Qwen 3.6 35B A3B
MTP… See the full description on the dataset page: https://huggingface.co/datasets/EXD-AI/vllm-configs.Qwen2.5-7B-Instruct-vllm-20251128_042753grammar-accuracy-qwen3.5-4b-trl-grpo-vllm-colocate-completions
TRL Completion logs
This dataset contains the completions generated during training using trl.
Find the trained model at https://huggingface.co/bihungba1101/grammar-accuracy-qwen3.5-4b-trl-grpo-vllm-colocate.
The completions are stored in parquet files, and each file contains the completions for a single step of training (depending on the logging_steps argument).
Each file contains the following columns:
step: the step of training
prompt: the prompt used to generate the completion… See the full description on the dataset page: https://huggingface.co/datasets/bihungba1101/grammar-accuracy-qwen3.5-4b-trl-grpo-vllm-colocate-completions.Nemotron-Personas-USA-synthetic-records-10files-qa-vllm-qwen4b-instruct-2507-clarqagumble-max-vllm-experimentvllm-metadatavllms-leaderboardprocessed_qwen25_7b_vllm_final_4processed_qwen25_7b_vllm_final_3deepseek-v4-flash-rocm-vllm-repro
Reproducing DeepSeek-V4-Flash on AMD ROCm with vLLM: 32K Correctness and TopK Sweep
This article summarizes an engineering reproduction of
deepseek-ai/DeepSeek-V4-Flash on an AMD ROCm ModelScope DSW instance. The work
focuses on a practical question: can a complex, fast-moving DeepSeek-V4-Flash
serving path be turned into a reproducible ROCm baseline with explicit
correctness gates?
The answer from this run is yes, with an important boundary: the current setup
is a fallback-heavy… See the full description on the dataset page: https://huggingface.co/datasets/lyydfys/deepseek-v4-flash-rocm-vllm-repro.output_3d_bounding_scannetppv2_vllm_old_descriptionprocessed_qwen25_7b_vllm_final
dataset_info:
features:
name: images
dtype: image
name: predictions
dtype: string
name: page_number
dtype: int64
name: title
dtype: string
name: author
dtype: string
name: thesis_id
dtype: string
name: university
dtype: string
name: department
dtype: string
name: year
dtype: string
name: language
dtype: string
name: thesis_type
dtype: string
name: keyword_abd
dtype: 'null'
name: abstract_tr
dtype: string
name: abstract_en
dtype: string
name: file_size_bytes
dtype: int64
name:… See the full description on the dataset page: https://huggingface.co/datasets/sghosts/processed_qwen25_7b_vllm_final.VLLM_ChartQA_splitprocessed_qwen25_7b_vllm_final_2Qwen2.5-7B-Instruct-vllm-retriever-20251202_093826res-vllm
ryzax/res-vllm
Copy of ryzax/res with vLLM-tool AIME25 runs swapped into the SUMM and LASTK folders. Original ryzax/res is unchanged.
All swapped/added runs are Qwen3-1.7B, AIME25 (30 problems × 64 samples), window 4096, max gen 262144, BUDGET_FORCE_SAVING=256 for summary variants.
Swaps
Path
What it is now
AIME25 avg@64
cov@64
maj@64
Qwen3-1.7B-SUMM/w4096s256
force256 + SAVING_PROMPT=context_system (evaluate-context example on last save only)
26.5%… See the full description on the dataset page: https://huggingface.co/datasets/prefixsliding/res-vllm.daily-paper-2026-07-21-agent-dynamic-batch-tuning-vllm
Dynamic Admission Reallocation for Multi-Tenant vLLM Serving
TL;DR — On a real H200 running vLLM, a dynamic controller that reallocates a fixed 96-slot admission budget toward live demand beat a frozen even 48/48 split by +14.0% total throughput (11,274 vs 9,894 tok/s) and +18.0% batch throughput (7,736 vs 6,557 tok/s) at equal-or-better p99 (4.08 vs 4.15 s). Honest caveat: steady-window GPU utilization reached only 67.2% mean (99% peak), short of a sustained 90% target… See the full description on the dataset page: https://huggingface.co/datasets/thaki-AI/daily-paper-2026-07-21-agent-dynamic-batch-tuning-vllm.modal-vllm-cache-b300-minimax-v44vLLM-SDF-and-SDF-plus-VT-rollouts-Qwen
vLLM SDF & SDF+VT rollouts Qwen
Full, untruncated Qwen3.6-35B-A3B rollouts generated locally with compiled CUDA graphs (vLLM 0.26.0 for the original exports and vLLM 0.19.1 CUDA 12.8 for the added SDF-2250 and VT-250 exports). The aggregate contains 21,290 records. Generation used temperature 0.3 and a 16,000-token per-turn/output cap.
Contents
Benchmark
Policies
Records
EvalAwareBench
VT-250, SDF-1250, SDF-2250, SDF-1250+VT-250, SDF-2250+VT-250
12,500… See the full description on the dataset page: https://huggingface.co/datasets/Usman391/vLLM-SDF-and-SDF-plus-VT-rollouts-Qwen.vLLM-SR-Preference-V1The files in this repo is the LLM-labeled samples that are used as the training dataset for vLLM-SR Preference model V1.
The training file (sharegpt_preference_labeld_with_negative.jsonl) contains 25k records that have sample_id, golden label for the preference-based routing policy, and a set of negative labels that are plausible but do not match the conversation context.
The validation file has the same structure, but only 1% of the training file size. The validation file and the training… See the full description on the dataset page: https://huggingface.co/datasets/ppppqp/vLLM-SR-Preference-V1.Nemotron-Personas-USA-synthetic-records-10files-qa-vllm-qwen4b-instruct-2507patch_tasks_vllm
Dataset Card for Patch-Based Visual Question Answering Dataset
Dataset Details
Dataset Description
This dataset contains approximately 305,000 triplets of question, answer, and image designed for patch-based visual reasoning tasks.
A standard question in this dataset is formatted as follows:
Image Grid: The image is divided into a 4x4 grid of 16 equal-sized patches. Patches are numbered sequentially from the top-left corner and moving right, then down to the… See the full description on the dataset page: https://huggingface.co/datasets/yurkes/patch_tasks_vllm.