datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
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.ToolMind
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.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.
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.Intermediate-Thinking-130k
Intermediate-Thinking-130k
A comprehensive dataset of 135,000 high-quality samples designed to advance language model reasoning capabilities through structured intermediate thinking processes. This dataset enables training and evaluation of models with sophisticated self-correction and iterative reasoning abilities across 42 languages.
OG Link
Overview
Intermediate-Thinking-130k addresses a fundamental limitation in current language models: their inability to pause… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/Intermediate-Thinking-130k.mlx-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.dolma3_mix-common_crawl-art_and_design-160kThe 160K subset of AllenAI's common_crawl-art_and_design Pretraining dataset split into train a valid saamples.
Train set size: 159436
Valid set size: 160
Direct usage in MLX-LM-LoRA:
python -m mlx_lm.lora \
--train \
--model Qwen/Qwen3-0.6B-Base \
--data mlx-community/dolma3_mix-common_crawl-art_and_design-160k \
--num-layers 4 \
--iters 1000 \
--batch-size 1 \
--steps-per-report 50 \
--max-seq-length 1028 \
--adapter-path path/to/adapter
Direct usage in MLX-LM:
python -m mlx_lm.lora \… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/dolma3_mix-common_crawl-art_and_design-160k.orpo-dpo-mix-40k-flat-mlx
Dataset Description
This dataset is a split version of orpo-dpo-mix-40k-flat for direct use with mlx-lm-lora, specifically tailored to be compatible with DPO and CPO training.
The dataset has been divided into three parts:
Train Set: 90%
Validation Set: 6%
Test Set: 4%
Example Usage
To train a model using this dataset, you can use the following command:
mlx_lm_lora.train \
--model Qwen/Qwen2.5-3B-Instruct \
--train \
--test \
--num-layers 8 \
--data… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/orpo-dpo-mix-40k-flat-mlx.MLX-Benchmark-V2
MLX Benchmark Dataset
Dataset Summary
The MLX Benchmark Dataset is a curated evaluation benchmark consisting of 520 questions designed to measure large language model (LLM) proficiency in Apple's MLX machine learning framework. MLX is an array framework for machine learning on Apple Silicon that leverages unified memory architecture, and this dataset is the first comprehensive benchmark specifically targeting MLX knowledge and coding ability.
The dataset covers the… See the full description on the dataset page: https://huggingface.co/datasets/Goekdeniz-Guelmez/MLX-Benchmark-V2.medfit-dataset
MEDFIT Medical QA Dataset
This dataset contains 6,444 unique healthcare-related question-answer pairs designed for fine-tuning language models for medical chatbot applications. The dataset was specifically curated for the MEDFIT-LLM research project focusing on domain-focused fine-tuning of small language models for healthcare applications.
All credits for the methodology and dataset creation go to Aditya Karnam Gururaj Rao, Arjun Jaggi, and Sonam Naidu.
The dataset has been… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/medfit-dataset.recycling_the_web-1m
Recycling the Web (MLX Subsets)
This is a subset of the facebook/recycling_the_web dataset, prepared for the MLX community.All credits for the original dataset go to Meta AI (Facebook).
Paper: Recycling the Web: A Method to Enhance Pre-training Data Quality and Quantity for Language Models
I’ve simply created smaller, more manageable shards for experimentation and training in MLX.Available sizes:
mlx-community/recycling_the_web-1k
mlx-community/recycling_the_web-100k… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/recycling_the_web-1m.JOSIE-DPO-Chosen-Ministral
JOSIE-DPO-Chosen — Ministral
Half of a DPO dataset. The chosen responses are here. You generate the rejected ones — and that's the point.
Overview
This dataset contains the chosen-only side of a preference dataset designed to align any LLM with the personality, tone, and response style of J.O.S.I.E. (Just One Super Intelligent Entity) — the viral model family created by Gökdeniz Gülmez.
The chosen responses were generated by a fine-tuned Ministral-14B model… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/JOSIE-DPO-Chosen-Ministral.JOSIE-Zero-8B-Reasoning-Traces-N67
JOSIE-Zero-Reasoning-Traces-N86
Reasoning traces generated by the JOSIE-ZERO-8B model.
JOSIE-ZERO-8B is a custom reasoning model trained using the GRPO (Group Relative Policy Optimization) training pipeline implemented in the MLX-LM-LoRA framework. The model was optimized with custom reward functions designed to encourage explicit reasoning, chain-of-thought style problem solving, self-correction, and structured analytical behavior.
This dataset contains high-quality reasoning… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/JOSIE-Zero-8B-Reasoning-Traces-N67.tnc-archive
Paraacademic institution's archived educational activities metadata scraped and packed as a dataset.
Dataset Details
Dataset Description
Scraped titles and summarized descriptions of the "non-members available" data of the Seminars of The New Centre for Research & Practice, took this from our website where i have a status of god of FireStoreStoNe (FSSN).
Regarding the latter, should i not scrap the data with an access to the direct descriptions links?
100%… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/tnc-archive.hallmark-mlx-reviewed-policy-traces
hallmark-mlx-reviewed-policy-traces
Reviewed citation-verification training traces for hallmark-mlx.
Contents
train.jsonl: 75 supervised examples
valid.jsonl: 6 supervised examples
Source reviewed traces: reviewed_seed_traces_combined.jsonl with 45 full traces.
Format
Each row is a prepared supervised training example for MLX LoRA fine-tuning.
The format is the exact snapshot used by the kept Qwen 1.5B run.
Upload Note
Review the… See the full description on the dataset page: https://huggingface.co/datasets/sebastianboehler/hallmark-mlx-reviewed-policy-traces.Dolci-Think-DPO-32B-FlatFlat version of AllenAI's Dolci-Think-DPO-32B.
Train set size: 199840
Valid set size: 160
MLX-LM-LoRA
mlx_lm_lora.train \
--model Goekdeniz-Guelmez/Josiefied-Qwen3-0.6B-abliterated-v1 \
--train \
--data mlx-community/Dolci-Think-DPO-32B-Flat \
--epochs 1 \
--batch-size 1 \
--num-layers 1 \
--val-batches 1 \
--steps-per-report 1 \
--adapter-path path/to/adapters \
--max-seq-length 1024 \
--grad-checkpoint \
--train-type lora \
--optimizer adamw \
--train-mode dpo \… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/Dolci-Think-DPO-32B-Flat.recycling_the_web-400K
Recycling the Web (MLX Subsets)
Paper: Recycling the Web: A Method to Enhance Pre-training Data Quality and Quantity for Language Models
This is a subset of the facebook/recycling_the_web dataset, prepared for the MLX community.All credits for the original dataset go to Meta AI (Facebook).
I’ve simply created smaller, more manageable shards for experimentation and training in MLX.Available sizes:
mlx-community/recycling_the_web-1k
mlx-community/recycling_the_web-100k… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/recycling_the_web-400K.dhanishtha-2.0-superthinker
📦 Dhanishtha-2.0-SUPERTHINKER-MLX
A distilled corpus of 11.7K high-quality samples showcasing multi-phase reasoning and structured emotional cognition. Sourced directly from the internal training data of Dhanishtha-2.0 — the world’s first Large Language Model (LLM) to implement Intermediate Thinking, featuring multiple <think> and <ser> blocks per response
Example with MLX-LM-LoRA:
mlx_lm_lora.train \
--model… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/dhanishtha-2.0-superthinker.orpo-dpo-mix-40k-mlx
Dataset Description
This dataset is a split version of orpo-dpo-mix-40k for direct use with mlx-lm-lora, specifically tailored to be compatible with ORPO training.
The dataset has been divided into three parts:
Train Set: 90%
Validation Set: 6%
Test Set: 4%
Example Usage
To train a model using this dataset, you can use the following command:
mlx_lm_lora.train \
--model Qwen/Qwen2.5-3B-Instruct \
--train \
--test \
--num-layers 8 \
--data… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/orpo-dpo-mix-40k-mlx.recycling_the_web-1k
Recycling the Web (MLX Subsets)
Paper: Recycling the Web: A Method to Enhance Pre-training Data Quality and Quantity for Language Models
This is a subset of the facebook/recycling_the_web dataset, prepared for the MLX community.All credits for the original dataset go to Meta AI (Facebook).
I’ve simply created smaller, more manageable shards for experimentation and training in MLX.Available sizes:
mlx-community/recycling_the_web-1k
mlx-community/recycling_the_web-100k… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/recycling_the_web-1k.Josiefied-Qwen3-dpo-v1-flatThis dataset has been used to create the first Josiefied beta version. This can be used directly within MLX-LM-LoRA.
Models used Qwen3-4B-4bit as the base (rejected) parts and Qwen3-30B-bf16 as the teacher (chosen) parts.
Usage MLX-LM-LoRA
mlx_lm_lora.train \
--model Goekdeniz-Guelmez/Josiefied-Qwen3-0.6B-abliterated-v1 \
--train \
--data mlx-community/Josiefied-Qwen3-dpo-v1-flat \
--epochs 1 \
--batch-size 1 \
--num-layers 1 \
--val-batches 1 \
--steps-per-report 1 \
--adapter-path… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/Josiefied-Qwen3-dpo-v1-flat.