datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
mlx-model-explorer-data
MLX Model Explorer Data
An anonymous record of how people use MLX Model Explorer to choose an MLX model for their Mac: which model families, sizes, quantizations, memory classes and context lengths they look at, and which models they go on to open, compare or download. It also holds the community reports ("it worked", "too slow") and real MLX benchmark results that people choose to contribute.
The goal is to answer, with data: what is the MLX community actually trying to run… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/mlx-model-explorer-data.optiq-lab-traces
OptiQ Lab Traces
Research and tool-calling sessions produced by OptiQ Lab, the local web UI that ships with mlx-optiq. Each session is a complete run: a deep-research report built from live web sources, or a multi-turn agent loop driving the Lab's own sandboxed tools.
The dataset is 866 sessions in HuggingFace Session-Traces format (the agent-traces viewer). Each .jsonl file is one session: a header line carrying the run's metadata, then one message per turn.
The two… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/optiq-lab-traces.optiq-code-traces
OptiQ Code Traces
Gold-verified agentic software-engineering trajectories, produced by OptiQ Code, the terminal coding agent for local models on a Mac. Each trajectory is a full tool-calling run against a real repository bug, and every resolved label is set by executing the gold tests (FAIL_TO_PASS + PASS_TO_PASS) after applying the model's patch, never by the agent's own self-report.
The dataset is 1,789 agent sessions in HuggingFace Session-Traces format (the agent-traces… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/optiq-code-traces.wikisql
mlx-community/wikisql
This dataset is processed from salesforce/WikiSQL and is used to demonstrate the fine-tuning function of LLM in the mlx_lm library.
source file form: mlx-examples
dots.mocr-mlx-evals
dots.mocr-mlx-evals
olmOCR-bench evals of dots.mocr MLX quants.
There is directory per model with the md files extracted from the original PDFs.
Directory logs contains the output from the harness run.
mlx-benchmarks
MLX Benchmarks
Structured benchmark results for MLX-quantized and other locally-hosted
LLMs on Apple Silicon. Covers throughput, time-to-first-token, tool-calling,
code generation, reasoning, knowledge, and math suites.
Results are produced by a sweep harness that wires upstream evaluation tools
against a local vllm-mlx inference server:
EleutherAI/lm-evaluation-harness — coding, reasoning, knowledge, math
linusvwe/MLXBench — throughput and time-to-first-token
vllm… See the full description on the dataset page: https://huggingface.co/datasets/JacobPEvans/mlx-benchmarks.dfn4_mlx_datastoreimagenet-1k
Dataset Summary
ILSVRC 2012, commonly known as 'ImageNet' is an image dataset organized according to the WordNet hierarchy. Each meaningful concept in WordNet, possibly described by multiple words or word phrases, is called a "synonym set" or "synset". There are more than 100,000 synsets in WordNet, majority of them are nouns (80,000+). ImageNet aims to provide on average 1000 images to illustrate each synset. Images of each concept are quality-controlled and human-annotated.
💡… See the full description on the dataset page: https://huggingface.co/datasets/mlx-vision/imagenet-1k.magenta-realtime-mlx-cpp
Magenta RealTime — C++ MLX runtime bundle
This dataset is a re-packaging of
Google's Magenta RealTime weights
for the C++ MLX runtime in
rhymeswithlion/magenta-realtime-mlx-cpp.
It contains exactly what mlx-stream needs at startup; nothing more, nothing
less. The upstream .pt / .npy checkpoints are intentionally not
mirrored here — they're only useful for the (Python) re-export tooling on the
project's main distribution.
Contents
.
├──… See the full description on the dataset page: https://huggingface.co/datasets/rhymeswithlion/magenta-realtime-mlx-cpp.Qwen3.8-27B-MLX-metricsToolMind
ToolMind: A Large-Scale, Reasoning-Enhanced Tool-Use Dataset
ToolMind is a large-scale, high-quality tool-agentic dataset with 160k synthetic data instances generated using over 20k tools and 200k augmented open-source data instances.
Our data synthesis pipeline first constructs a function graph based on parameter correlations and then uses a multi-agent framework to simulate realistic user–assistant–tool interactions.
Beyond trajectory-level validation, we employ fine-grained… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/ToolMind.mlx-omni-lora-stt-tts-demoWill be used in the development of the trainer backend of mlx-omni by Neywa Labs.
seeclick-web-commercial-mlx
SeeClick Web Commercial Dataset (MLX-VLM Format)
Commercial-use friendly GUI grounding dataset from SeeClick Web data.
Apache 2.0 licensed - safe for commercial applications.
Dataset Description
This dataset contains ~20k examples for training Vision-Language Models to predict
click coordinates given a screenshot and instruction. Derived from SeeClick Web
crawled data (Apache 2.0).
Key Features
License: Apache 2.0 (commercial use allowed)
Format: MLX-VLM… See the full description on the dataset page: https://huggingface.co/datasets/pierretokns/seeclick-web-commercial-mlx.iclr2027_univtac_mlx_ckptqwen3_dwq_calibration_1332_235b
Qwen3 DWQ Calibration Dataset (235B, 1332 samples)
This dataset contains 1,332 samples for calibrating dynamic weight quantization (DWQ) of Qwen3-235B models. It is created following the methodology of mlx-community/qwen3_dwq_calibration_1332 but using the larger Qwen3-235B model.
Dataset Description
This dataset is derived from allenai/tulu-3-sft-mixture and consists of:
610 samples processed through Qwen3-235B with explicit reasoning
722 original samples from the… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/qwen3_dwq_calibration_1332_235b.JOSIE-v2-Instruct-5K
JOSIE v2 Instruct 5K
A high-quality instruction-following dataset featuring J.O.S.I.E. (Just One Super Intelligent Entity) - an advanced AI assistant with a distinctive personality combining intellectual rigor, dry wit, and genuine helpfulness.
Dataset Overview
Size: 5,000 conversational samples
Format: JSONL (JSON Lines)
Source Model: GPT-5.4-nano via OpenAI Batch API
Use Case: Finetuning language models on Apple Silicon using mlx-lm or mlx-lm-lora
License: MIT… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/JOSIE-v2-Instruct-5K.Apertus-v1.5-QAT-10K
mlx-community/Apertus-v1.5-QAT-10K
This is a 2000 sample subset of the chosen pairs inside swiss-ai/Apertus_v1p5_Preference_Data for MLX-LM-LoRA and MLX-LoRA-Studio and the Quantization Aware Trained Appertus models.
hermes-reasoning-tool-usefineweb-edu-1m
train splitt = 1M
validation nsplitt = 1024
traces.claude-code.mlx-lm-granitemoehybridmlx7-two-tower-data
mlx7-two-tower-data
This repository contains datasets used for training Two-Tower (Dual Encoder) models for document retrieval.
Dataset Description
The datasets provided here are structured for training dual encoder models with various sampling strategies:
classic_triplets: 48.2 MB
intra_query_neg: 47.6 MB
multi_pos_multi_neg: 126.5 MB
Dataset Details
classic_triplets.parquet: Standard triplet format with (query, positive_document, negative_document)… See the full description on the dataset page: https://huggingface.co/datasets/Azuremis/mlx7-two-tower-data.openhermes-2.5-mlxLiterally just https://huggingface.co/datasets/teknium/OpenHermes-2.5 but converted to be usable in MLX lora training (assumes ChatML format)
fineweb-edu-100KCode used to create this:
import json
from itertools import islice
from datasets import load_dataset
from tqdm import tqdm
DATASET_NAME = "HuggingFaceFW/fineweb-edu"
DATASET_CONFIG = "sample-10BT"
NUM_SAMPLES = 100_000
BATCH_SIZE = 5_000
OUTPUT_PATH = "pretrain_100k.jsonl"
dataset = load_dataset(
DATASET_NAME,
DATASET_CONFIG,
split="train",
streaming=True,
)
dataset_iter = iter(dataset)
written = 0
with open(OUTPUT_PATH, "w", encoding="utf-8") as f:
progress = tqdm(… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/fineweb-edu-100K.mlx-serve-macos-15-build-run
mlx-serve on macOS 15.x — build & run package (Qwen3.8-Flash-Next with MTP)
Proven 2026-09-05 on macOS 15.7.7, M2 Max 96 GB, Apple clang 17
(CommandLineTools 16.4, no full Xcode at start) → working
mlx-serve 26.9.1-dev serving Qwen3.8-Flash-Next (qwen4_exp) 4-bit pack
with native MTP at TG 27.43 t/s / PP 29.6 t/s.
Model pack source: https://huggingface.co/ddalcu/Qwen3.8-Flash-Next-MLX-Serve-4bit
mlx-serve upstream: https://github.com/ddalcu/mlx-serve/ (site: mlxserve.com)
This… See the full description on the dataset page: https://huggingface.co/datasets/ljupco/mlx-serve-macos-15-build-run.vllm-wheelsxlam-function-calling-60k-mlxmlx-local-inference-benchmarks
MLX local-inference benchmarks — Qwen3.6 & Laguna-S/XS families
Raw results, harnesses and methodology for an 8-axis benchmark of four MLX
checkpoints on a 128 GB M5 Max. Everything a person would need to check my numbers
or disagree with them.
Companion model repos:
Tess-4-27B-MLX-Q8 — with a working MTP head
Tess-4-27B-MLX-Q4 — same, at 4-bit
NEW (2026-07-24): the Laguna chapter — REPORT-LAGUNA.md + results-laguna/
Five-way same-engine bake-off (Laguna-S… See the full description on the dataset page: https://huggingface.co/datasets/studioburnside/mlx-local-inference-benchmarks.medical-dataset-test-mlxgsm8kOpenAI's GSM8K dataset converted to be compatibel with MLX-LM-LoRA.
example uasge:
pip install -U mlx-lm-lora
python -m mlx_lm_lora.train \
--model mlx-community/Josiefied-Qwen3-0.6B-abliterated-v1-4bit \
--train \
--train-mode grpo \
--data mlx-community/gsm8k \
--iters 100 \
--steps-per-report 1 \
--batch-size 1 \
--max-completion-length 512
swift-mlx-Qwen3-Embedding-4B
🔍 VincentGOURBIN/swift-mlx-Qwen3-Embedding-4B - Embeddings Dataset
Description
Ce dataset contient des embeddings vectoriels générés par le système LocalRAG pour la recherche sémantique dans la documentation technique.
📊 Statistiques
Format: SafeTensors
Vecteurs: 7,511
Dimension: 2560
Modèle d'embedding: Qwen/Qwen3-Embedding-4B
Type d'index: HNSW
Généré le: 2025-08-22T14:04:16.932676
📁 Contenu
embeddings.safetensors: Embeddings vectoriels au… See the full description on the dataset page: https://huggingface.co/datasets/VincentGOURBIN/swift-mlx-Qwen3-Embedding-4B.