datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
mtpnet_image_models
模型训练过程汇总[该仓库只含有image model的训练过程]
本仓库采用扁平化的目录结构和标签系统来组织模型，具体说明如下：
仓库结构
一级目录：直接以模型名称-数据集，例如 ResNet-CIFAR-10、GraphMAE_QM9-Cora 等
二级目录：包含该模型在该数据集下的不同训练任务或变体，例如 normal、noisy、backdoor_invisible 等
训练过程目录结构：每个模型目录下包含：
scripts/：存放模型相关代码和训练脚本
epochs/：存放模型训练过程和权重文件
每个epoch的权重文件(model.pth)和embedding(.npy)
dataset/：模型需要的数据集
仓库结构展示
文件结构展示
mtpnet_tokens
模型训练过程汇总（持续更新中）
对于已收集的每一个模型，code 目录为模型定义、训练和测试的代码和脚本文件，model 目录为已收集的 epoch 模型文件，dataset.zip 为模型数据集。
下表汇总了所有收集的模型训练过程信息：
模型名称
模型简介
模型类型
Epoch数量
数据集信息
Clone-detection-BigCloneBench
基于大规模代码克隆基准数据集的代码克隆检测模型，任务是进行二元分类(0/1),其中1代表语义等价，0代表其他情况。
代码克隆检测
2个epoch
BigCloneBench数据集
Clone-detection-POJ-104
基于POJ-104数据集的代码克隆检测模型，任务是识别不同编程题目中相似的代码实现，给定一段代码和一组候选代码，任务是返回具有相同语义的Top K个代码
代码克隆检测
2个epoch (0-1)
POJ-104编程题目数据集… See the full description on the dataset page: https://huggingface.co/datasets/code-philia/mtpnet_tokens.qwen36-kquant-offload-mtp-swebench-lite100-results
Qwen3.6 K-Quant Offload MTP SWE-bench Lite 100 Results
This dataset contains the complete 5-model x 100-prompt runtime benchmark artifacts plus a detailed statistical analysis layer.
Primary conclusion: hot30/cold30 was the best decode-throughput run, while Q4_K_M had the best total wall clock. The ATX hot30/cold30 quantization significantly outperformed both Q4_K_M and Q3_K_XL on paired decode throughput, but Q4_K_M remains the elapsed-time control.
The ATX/K3 hot10, hot20, and… See the full description on the dataset page: https://huggingface.co/datasets/jakeatx/qwen36-kquant-offload-mtp-swebench-lite100-results.mtpnet_graph_models
模型训练过程汇总[该仓库只含有graph model的训练过程]
本仓库采用扁平化的目录结构和标签系统来组织模型，具体说明如下：
仓库结构
一级目录：直接以模型名称-数据集，例如 ResNet-CIFAR-10、GraphMAE_QM9-Cora 等
二级目录：包含该模型在该数据集下的不同训练任务或变体，例如 normal、noisy、backdoor_invisible 等
训练过程目录结构：每个模型目录下包含：
scripts/：存放模型相关代码和训练脚本
epochs/：存放模型训练过程和权重文件
每个epoch的权重文件(model.pth)和embedding(.npy)
dataset/：模型需要的数据集
仓库结构展示
文件结构展示
bonsai2-27b-mtp-repro
Ternary-Bonsai-2-27B + in-file MTP: reproduction bundle (RTX 4080 SUPER, Ada/SM89)
This repository holds the raw data. The method (build script, harness, launch units, write-up) lives on GitHub:
https://github.com/zhaoyilun/bonsai2-27b-mtp-repro
Both are the same piece of work: the GitHub repo has the code and the how-to, this dataset has the
measurements it produced. Cross-linked in both directions.
Raw measurements, scripts and notes for the two discussions:
official model… See the full description on the dataset page: https://huggingface.co/datasets/zhaokeqi/bonsai2-27b-mtp-repro.mt_pubmedMTPmtp-selfdata-qwen3-8b-finewikiqwen36-mtp-turbo-kv-analysis
Qwen3.6 MTP Turbo KV Runtime Analysis
This repository is a curated analysis artifact for local Qwen3.6-35B-A3B MTP GGUF inference experiments on Windows CUDA. It compares clean MTP llama.cpp, QuinsZouls llama-next TurboQuant, and the completed subset of Atomic TurboQuant runs under a fixed 64k context, MoE CPU offload, and Unsloth-aligned sampling settings.
The raw benchmark runs included incomplete and capability-incompatible rows. This repo keeps only completed, comparable… See the full description on the dataset page: https://huggingface.co/datasets/jakeatx/qwen36-mtp-turbo-kv-analysis.kimi-mtp-dataset
Kimi-K2.5 Eagle3 Training Data
This dataset contains the instruction-following data used to train an Eagle3 MTP draft model for Kimi-K2.5 with TorchSpec.
All responses were regenerated by running Kimi-K2.5 via Engine rather than taken from the original datasets. This is critical for speculative decoding training: the draft model must learn the exact token-level distribution of the target model it is accelerating.
The trained Eagle3 draft model is available at… See the full description on the dataset page: https://huggingface.co/datasets/lightseekorg/kimi-mtp-dataset.mtp-selfdata-qwen3-8b-finewiki-40kqwen36-27b-mtp-long-context-decay
Qwen3.6-27B MTP Long-Context Decay Benchmark
This artifact contains a local Apple Silicon benchmark of Unsloth Qwen3.6-27B GGUF quantizations running with llama.cpp MTP draft-2 speculative decoding. The benchmark measured generation speed over sequential 1K-token windows up to 16K generated tokens, across context caps and KV cache precision.
Hub repo: sjakek/qwen36-27b-mtp-long-context-decay
Generated locally: 2026-05-14T08:51:11Run directory on source machine:… See the full description on the dataset page: https://huggingface.co/datasets/jakeatx/qwen36-27b-mtp-long-context-decay.glm-dsa-mtp-audit
GLM-5.2 (GLM_DSA) NextN/MTP for llama.cpp — validation & wiring audit
Reference evidence for the llama.cpp pull request adding NextN / multi-token-prediction (MTP) speculative decoding for the GLM_DSA architecture (GLM-5.2) as a --spec-type draft-mtp target.
This repository exists so reviewers can inspect the raw correctness and performance evidence directly, rather than taking summary numbers on faith. It contains: the operation-by-operation wiring audit against the vLLM… See the full description on the dataset page: https://huggingface.co/datasets/satgeze/glm-dsa-mtp-audit.bge-mtp-datagemma4-mtp-fixturesmtp-selfdata-llama3.2-3b-finewikinvfp4-mtp-survey
Do Qwen3.8-27B NVFP4 repos actually ship a working MTP draft head?
A static survey of every NVFP4 quantization of Qwen3.8-27B and its finetunes that I could find
on the Hugging Face Hub, last run on 2026-08-24 (Rev 4) with
nvfp4_mtp_audit.py. Raw output: results.json.
I ran this to check a claim I had made in public, and the claim did not survive. The correction
is the first section, because it is the most important result here.
Revision history — read this, it is… See the full description on the dataset page: https://huggingface.co/datasets/windowsxp811203/nvfp4-mtp-survey.pr3118-mtp-e2e-validation
PR3118 MTP validation
19 active model repositories, matching the PR metrics table: seven earlier Qwen dense variants plus three Qwen MTP pipeline checkpoints, one Qwen MoE preservation variant, six Nemotron variants, and the full pretrained GLM-5.3 Flash and DSA checkpoints.
Full-model MTP serving smoke results:
GLM-5.3 (DSA): 58/70 draft tokens accepted (82.86%); one 128-token completion, four B200s, TP=4, vLLM 0.28.0, one speculative MTP token. Request, response, and… See the full description on the dataset page: https://huggingface.co/datasets/soyrsoyr/pr3118-mtp-e2e-validation.GLM-5.2-MTP78-calibration-capture
GLM-5.2 MTP Layer-78 Calibration Capture
7,288,310 tokens of layer-78 (MTP draft) MoE calibration data for
GLM-5.2 753B, captured during live MTP serving on the
GLM-5.2-EXL3-TR3-3.0bpw
checkpoint: every token's MoE input hidden state, paired with the router's
ground-truth top-8 expert assignment.
With this dataset, re-encoding layer 78 at any bitrate is a pure-compute
job — no 4-GPU capture rig, no serving stack, no corpus drive needed. It
calibrated the published… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/GLM-5.2-MTP78-calibration-capture.gemma4-mtp-quality-study
Gemma 4 MTP Quality–Throughput Study
Author: Taiko Toeda, Independent ResearcherORCID: https://orcid.org/0009-0001-7267-0201Version: 1.0.0DOI: 10.5281/zenodo.21860461
This Dataset repository distributes the sanitized processed measurements from
an independent paired benchmark of QAT Q4_0 Gemma 4 12B IT and 26B-A4B IT on
one RTX 5070 Ti using pinned CUDA llama.cpp commit
7ba604f1cb61cd14898138e9abc0b4ff2601f180.
Included data
request-level timing, throughput… See the full description on the dataset page: https://huggingface.co/datasets/moebiusT7/gemma4-mtp-quality-study.vizdoom_hgs_fullredhat-docs_dataset
🖥️ Red Hat Technical Documentation Dataset
📌 Overview
This dataset contains 55,741 structured technical documentation entries sourced from Red Hat, covering:✅ System Administration Guides – User management, permissions, kernel tuning✅ Networking & Security – Firewall rules, SELinux, VPN setup✅ Virtualization & Containers – KVM, Podman, OpenShift, Kubernetes✅ Enterprise Software Documentation – RHEL, Ansible, Satellite, OpenStack
📊 Dataset Details
This… See the full description on the dataset page: https://huggingface.co/datasets/mtpti5iD/redhat-docs_dataset.glm52-fidelity-exl3-tr3v4-3.5bpw-mtp78-brandonmusic-v1
fidelity--glm52.malaiwah.quant.exl3-tr3v4-3.5bpw-mtp78-brandonmusic
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from brandonmusic/GLM-5.2-EXL3-TR3v4-3.5bpw-MTP78.
The cut
the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm52-fidelity-exl3-tr3v4-3.5bpw-mtp78-brandonmusic-v1.mt_puzzles
Multi-Turn Puzzles: Evaluating Interactive Reasoning and Strategic Dialogue in LLMs
MT-Puzzles is a novel benchmark comprising a suite of multi-turn tasks each designed to test specific reasoning, interactive dialogue, and information-seeking abilities:
Word Guess Guess the secret word in min attempts while environment gives feedback on how close the guess is at each turn.
Movie Recommendation: Probe the user to decode the user preference function for N turns. Pick a movie for the… See the full description on the dataset page: https://huggingface.co/datasets/arianhosseini/mt_puzzles.mtp-selfdata-qwen3.6-35b-a3b-finewikirecap_dish_20260611This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "bi_so101_follower",
"total_episodes": 71,
"total_frames": 39618,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 500,
"fps": 30,
"splits": {
"train": "0:71"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/mt-prox/recap_dish_20260611.qwodasuce-mtp-v2MT-prefrecap_intervention0625This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "bi_so101_follower",
"total_episodes": 42,
"total_frames": 64742,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 500,
"fps": 30,
"splits": {
"train": "0:42"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/mt-prox/recap_intervention0625.nexus-sft-mix-v6-phase2-mtprce
NEXUS SFT Mix v6 Phase 2 — MTP-RCE Option B
149,799 samples for NEXUS phase 2 continuation training. Includes MTP-RCE mode annotations on 16% of samples.
Composition
105,000 rehearsal v5 : stratified subset of mix v5 (50 sources, all preserved for anti-forgetting). Preserves identity (4,730 nexus_identity_v2) + style + bilingue 60% FR / 40% EN.
20,799 HF reasoning : math/reasoning filtered ≤4000 chars (NEXUS seq_len 4096 safe).
12K… See the full description on the dataset page: https://huggingface.co/datasets/jescy525/nexus-sft-mix-v6-phase2-mtprce.