datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
DeepSeek-V4.1-Flash-NVFP4-metrics
DeepSeek-V4.1-Flash-NVFP4 metrics
Everything behind the numbers in AtomicChat/DeepSeek-V4.1-Flash-NVFP4-nvidia.
logprobs/lp-<run>-<corpus>.npz: the raw top-512 log probabilities of every measurement run, 49,152 scored
positions each: ref, ref-repeat, ref-r3, ref-b1 (batch size 1) for the original; flat, flat-r2,
flat-r3 for the uncalibrated cast; nvidia, nvidia-r2, nvidia-r3 for the calibrated checkpoint.
logs/kld-<run>-<corpus>.json: the KL lower bound per run against ref… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/DeepSeek-V4.1-Flash-NVFP4-metrics.glm52-fidelity-nvfp4-nvidia-v1
fidelity--glm52.malaiwah.quant.nvfp4-nvidia
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from nvidia/GLM-5.2-NVFP4.
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 time: the capture already sits after it). Same cut… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm52-fidelity-nvfp4-nvidia-v1.nemotron-super-120b-nvfp4-imo-answerbenchglm-5.2-nvfp4-agentic-traces
GLM-5.2 NVFP4 agentic software traces
This snapshot contains 1,989 completed Verifiers invocation records generated
with RedHatAI/GLM-5.2-NVFP4-FP8.
manifest.jsonl is a compact index for filtering and inspection.
data/<arm>.jsonl contains the exact full graph records emitted by Verifiers.
configs/<arm>.toml contains the resolved configuration for each arm.
The snapshot retains successes, failures, truncations, and scoring metadata.
Use solved, reward, has_error, failure_labels… See the full description on the dataset page: https://huggingface.co/datasets/synquid/glm-5.2-nvfp4-agentic-traces.minimax-h3-nvfp4-data
MiniMax-H3 NVFP4 Data
Large data for the NVFP4 static sparse attention experiment (2026-09-11).
Project entry point: H3 Attention Lab.
The project README contains result tables, audit limitations, source code, small evidence files, and public download instructions.
This revision contains 84 data files (752,104,680 bytes):
70 generated videos: experiment/outputs/*/*/quality.mp4
3 comparison videos: experiment/comparisons/*__e2e_comparison.mp4
10 reference videos:… See the full description on the dataset page: https://huggingface.co/datasets/Zer0-Sky/minimax-h3-nvfp4-data.Inkling-Small-NVFP4-Regenerated-Collection
Inkling-Small-NVFP4 Regenerated Collection
On-policy training data for a DSpark
speculative-decoding drafter targeting
thinkingmachines/Inkling-Small-NVFP4.
Every assistant response here was regenerated by Inkling-Small-NVFP4 itself over prompts
drawn from Magpie + UltraChat, so the completions reflect the target model's own
distribution rather than the datasets' original responses. This is what makes the data
on-policy for drafter training: the drafter learns to predict the… See the full description on the dataset page: https://huggingface.co/datasets/orestis-z/Inkling-Small-NVFP4-Regenerated-Collection.daily-paper-2026-07-12-nvfp4-moe-selective-quant
Traffic-Aware Selective NVFP4 Quantization for MoE LLMs
TL;DR — On a real decoder MoE (allenai/OLMoE-1B-7B-0924, 1024 experts, top-8) with a valid perplexity metric, traffic-aware selective 4-bit quantization works: protecting the highest-traffic (hot) experts at full precision and quantizing the rest Pareto-dominates a same-storage random (linear-mix) allocation at every budget, recovering 46.9% of the uniform-to-bf16 perplexity gap at a ~24% storage premium (hottest 10%) and… See the full description on the dataset page: https://huggingface.co/datasets/thaki-AI/daily-paper-2026-07-12-nvfp4-moe-selective-quant.dgx-spark-nvfp4-notes
dgx-spark-nvfp4-notes
Assets for the Hugging Face discussion on nvidia/Qwen3.6-35B-A3B-NVFP4: official DGX Spark Marlin recipe dies under concurrent json_schema load on GB10 / SM121.
nvfp4-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.calib-agentic-sample-Qwen3.8-27B-QUASAR-NVFP4
calib-agentic-sample — Qwen3.8-27B-QUASAR-NVFP4
The exact calibration sample used to quantize lm_head in
digi-texx/Qwen3.8-27B-FULL-NVFP4.
Published so the quantization is reproducible: this is not a representative extract, it is the
documents the quantizer actually saw.
Lineage
11 public agentic / tool-calling datasets
-> digi-texx/calib-agentic-normalized 6,044,537 rows (unified schema)
-> digi-texx/calib-agentic-curated 5,835,723 rows… See the full description on the dataset page: https://huggingface.co/datasets/digi-texx/calib-agentic-sample-Qwen3.8-27B-QUASAR-NVFP4.glm53-fidelity-nvfp4-inferact-v1
fidelity--glm53.malaiwah.quant.nvfp4-inferact
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from Inferact/GLM-5.3-NVFP4.
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 time: the capture already sits after it). Same… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm53-fidelity-nvfp4-inferact-v1.glm53-fidelity-nvfp4-radixark-v1
fidelity--glm53.malaiwah.quant.nvfp4-radixark
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from RadixArk/GLM-5.3-NVFP4.
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 time: the capture already sits after it). Same… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm53-fidelity-nvfp4-radixark-v1.glm53-fidelity-nvfp4-incoai-v1
fidelity--glm53.malaiwah.quant.nvfp4-incoai
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from incoai/GLM-5.3-NVFP4.
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 time: the capture already sits after it). Same cut… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm53-fidelity-nvfp4-incoai-v1.gpt-oss-20b-NVFP4fleetwide-nvfp4-t4-wheelhouse
fleetwide-nvfp4 T4 wheelhouse
Prebuilt wheels for the Google Colab T4 / SM75 path.
Target
Ubuntu 22.04
glibc 2.35
Python 3.12 / cp312
CUDA 13.0
SM75 / T4
vLLM ref: fleetwide-nvfp4
vLLM commit: 5afe8787341b8b39ad0aa81d5f0331704cfa4c11
Artifact
URL:
https://huggingface.co/datasets/mgschwind/fleetwide-nvfp4-t4-wheelhouse/resolve/main/fleetwide_nvfp4_t4_wheelhouse.tgz
SHA256:
40d4493dbd12d251c01f24ddff9b9a5b2b01b680ed2c22fc1c3604caa04a3696
Contents:… See the full description on the dataset page: https://huggingface.co/datasets/mgschwind/fleetwide-nvfp4-t4-wheelhouse.dataset-biased-nvfp4-18Kvideos_vbench_5b_long30s_nvfp4nvfp4_videosIDA-blackwell-nvfp4-receipts
IDA — Blackwell NVFP4 target-GPU receipts
Target-GPU evidence for the IDA-TRAIN-V2 native engine's staged NVFP4 path
(precision_profile=blackwell_nvfp4), on open-weight model shapes only
(Qwen2.5, SmolLM2, GPT-2). These are evidence bundles, not model weights and
not promotion receipts — every run is profile_only / mlperf_canary,
carries placeholder-weight checkpoints, and is not promotion_eligible. Read
them alongside docs/nvfp4-rollout.md in the source repo (the "Promotion… See the full description on the dataset page: https://huggingface.co/datasets/KissTheHabit/IDA-blackwell-nvfp4-receipts.movie_gen_60s_real_refine_step4_nvfp4_KVterminal-bench-nvfp4-opencode-runs
Terminal-Bench 2 OpenCode NVFP4 Runs
This dataset contains snapshots from Terminal-Bench 2 runs using Margin Lab evals, OpenCode, vLLM, and NVFP4 model checkpoints. The runs were stopped before completion after the captured artifacts were synced and uploaded.
The canonical snapshots are under:
mac/terminal_bench_nvfp4_results/
Older direct node uploads also exist under brev/; use the mac/ tree for the final synced state.
Runs
Run path
Model
Status… See the full description on the dataset page: https://huggingface.co/datasets/dmcc73/terminal-bench-nvfp4-opencode-runs.movie_gen_30s_real_refine_step4_nvfp4noun-courseware-nvfp4-256toksamantha-nvfp4-calibrationrobot_demo_nvfp4movie_gen_60s_real_refine_step4_nvfp4video_demo_WALLE_nvfp4noun-courseware-nvfp4-1024tokstallion-backup-w2v-bert-nvfp4-20260713rag_qwen3_32b_nvfp4_results