datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
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.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.gemma4-mtp-fixturesnvfp4-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.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.ATOM_regen_seeklight_kimi_mtp
ATOM regen: seeklight kimi-mtp responses by Kimi-K3
用 Kimi-K3 对
lightseekorg/kimi-mtp-dataset
的 prompt 重新生成了一遍回答，供 off-policy 投机解码蒸馏（SDDD）使用。
原始 pipeline 每轮都要用 teacher 重新解码一次（Phase A1）。把回答预生成并缓存下来，
A1 整个消失，之后每一轮训练都直接复用，代价从「每轮一次」变成「一共一次」。
数据
450,625 行，每行一段对话：
{"conversations": [
{"role": "user", "content": "..."},
{"role": "assistant", "reasoning_content": "...", "content": "..."}
]}
reasoning_content 是 K3 的 thinking 内容，和 content 分开存。
多轮对话保留了历史轮次里完整的 assistant… See the full description on the dataset page: https://huggingface.co/datasets/slippedJim/ATOM_regen_seeklight_kimi_mtp.kimi-k2.5-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 SGLang 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/UnicornChan/kimi-k2.5-mtp-dataset.hy3-w4a16-mtp-calibration
Hy3 W4A16-MTP — Calibration Set
The exact 512-sample calibration blend used to GPTQ-quantize
canada-quant/hy3-w4a16-mtp
(a W4A16 quantization of tencent/Hy3).
Published for full reproducibility of the quantization pipeline.
Why a blend (not chat-only)
INT4 weight quantization degrades most on code and tool-call-shaped tokens. A chat-only
calibration set (e.g. pure ultrachat) under-samples exactly the routed experts those tokens
activate. This set deliberately… See the full description on the dataset page: https://huggingface.co/datasets/canada-quant/hy3-w4a16-mtp-calibration.MTP-finetune-datamtp