datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Qwen3.8-27B-GGUF-metrics
Qwen3.8-27B GGUF, everything behind the numbers
This is the working record for
AtomicChat/Qwen3.8-27B-GGUF.
Every figure in that model card came from a file in here, including the ones
about other publishers' builds.
The point of publishing it is simple. A quantization comparison is only worth
reading if someone else can run it, and that needs three things nobody usually
ships: the exact reference the numbers were measured against, the exact text
they were measured on, and the… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/Qwen3.8-27B-GGUF-metrics.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.Muse-Glimmer-30B-GGUF-metrics
Muse Glimmer 30B GGUF — raw metrics
Every log behind the numbers in
AtomicChat/Muse-Glimmer-30B-GGUF.
Published unfiltered, so any figure in the model card can be checked or disputed.
Layout
Path
Contents
kld/
llama-perplexity --kl-divergence output, per build and per corpus
bench/
llama-bench -o json
speculative/
llama-server logs with and without the drafter
layouts/
per-tensor type map of every GGUF
conversion/
convert_hf_to_gguf.py logs… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/Muse-Glimmer-30B-GGUF-metrics.Ornith-1.5-35B-A3B-GGUF-metricscalib-corpora
calib-corpora
A pool of calibration material, the recipes that turn it into a calibration set
for one specific model, and the measurement corpora those quants are scored
against.
This repository is not a corpus. Nothing here is meant to be fed to
llama-imatrix as-is except the files under builds/, and each of those was
made for one named model and is close to useless for any other.
Why it is built this way
The first version of this repository was a single… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/calib-corpora.dsv4-eval-artifacts
DeepSeek-V4-Flash-0731 — quantization measurements
Everything needed to reproduce, audit or extend the numbers published in
AtomicChat/DeepSeek-V4-Flash-0731-GGUF:
the reference logits, the evaluation corpus, the raw tool output for every quant we
measured, and the parsed results.
Every GGUF of this model that we could find on the Hub was measured here — ours,
unsloth's, bartowski's, ggml-org's, antirez's and others — on one machine, against one
reference, with one command.… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/dsv4-eval-artifacts.Qwen-Image-2.1-Turbo-Abliterated-Uncensored-GGUF-metricsembeddinggemma-2-GGUF-metrics
embeddinggemma-2 GGUF, everything behind the numbers
This dataset holds the measurements, logs and inputs behind
AtomicChat/embeddinggemma-2-GGUF.
What is here
Path
What it is
results.json
Every measured file: ours, Unsloth's, AutoRound's (webmp3/Sakura-EmbeddingGemma-2-AutoRound-GGUF) and ggml-org's Q8_0. Each row has the size and, per eval set and width (768, 256), the mean and p99 of 1 - cosine to BF16, the same-top-result rate with its 95% interval… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/embeddinggemma-2-GGUF-metrics.Ling-3.0-flash-GGUF-metrics
Ling-3.0-flash — quantization metrics
Everything measured while building the GGUF line for inclusionAI/Ling-3.0-flash: raw logs, per-rung numbers and the importance matrix statistics. Published so the quant table can be checked rather than trusted.
Quants live in AtomicChat/Ling-3.0-flash-GGUF.
Layout
metrics/
grid-table.json per rung: size, bpw, mean/99% KLD, top-1 agreement
kld-results.json raw parser output of every KL divergence run… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/Ling-3.0-flash-GGUF-metrics.Qwen3.8-Flash-Next-GGUF-metricsd1-omni-600M-GGUF-metrics
d1-omni-600M-GGUF metrics
The measurements and inputs behind AtomicChat/d1-omni-600M-GGUF.
results.json: every file against the original weights (FP32, run through Liquid's PyTorch code). It lists
the changed answers, the same-answer rate, the mean and p99 option KL, and the mean and largest option total
variation distance.
decisions/: the option probabilities for each file and decision. reference.json comes from the original
weights, and liquid-Q8_0.json from Liquid's own Q8_0… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/d1-omni-600M-GGUF-metrics.d1-3B-GGUF-metrics
d1-3B GGUF, everything behind the numbers
These are the measurements and inputs behind
AtomicChat/d1-3B-GGUF.
Path
What it is
results.json
Every measured file, ours and Liquid's, with its size and two sets of numbers. Decision fidelity against BF16: same answer, flips, option KL and total variation distance. Text: KLD and top-1 over eval/neutral
imatrix/imatrix.gguf, imatrix/calib.txt
The importance matrix and its 1.5M-token corpus of decision prompts… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/d1-3B-GGUF-metrics.Qwen3.8-Flash-Next-Abliterated-Uncensored-GGUF-metrics
Qwen3.8-Flash-Next-Abliterated-Uncensored-GGUF: measurements
Everything behind the numbers on the
model card,
from one run on 2026-10-07/08: Qwen/Qwen3.8-Flash-Next@de4b8e4d, llama.cpp
980aef8c, 8x RTX PRO 6000 (sm_120), CUDA 13.
Path
What
kld/
The original BF16 model's logits over the held-out neutral and code sets (87 chunks at 4096 context), the reference for every KLD.
ablit/data/manifest.json
Prompt sources with revisions, the split sizes and the sha256 of… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/Qwen3.8-Flash-Next-Abliterated-Uncensored-GGUF-metrics.Ornith-1.5-9B-GGUF-metricsQwen3.8-27B-MLX-metricsTernary-Bonsai-2-27B-Abliterate-LoRA-GGUF-metrics
Ternary-Bonsai-2-27B-Abliterate-LoRA-GGUF-metrics
Measurements behind
AtomicChat/Ternary-Bonsai-2-27B-Abliterate-LoRA-GGUF,
the rank-1 refusal-ablation adapter for PrismML's 1.75 bit/weight ternary pack.
Both packs are covered: every row carries a pack column, PTQ1_0 or PQ2_0. The same
adapter file was run on both, and on the refusal evaluation all 416 greedy replies came out
byte-identical across packs.
Aggregates only. Prompt text is not redistributed (the sources are named in… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/Ternary-Bonsai-2-27B-Abliterate-LoRA-GGUF-metrics.Qwen-Image-2.1-Turbo-GGUF-metrics
Qwen-Image-2.1-Turbo GGUF metrics
Everything behind the numbers in AtomicChat/Qwen-Image-2.1-Turbo-GGUF,
including the ones about other publishers' files. Every render in the tables is
here, so any number can be recomputed, and any quant, ours or not, can be
measured against exactly the reference we used.
path
what
images/ref/
the reference: the bf16 denoiser, 1024x1024 steps=8 cfg=1.0 euler sigmas=[1.0,0.978453,0.95418,0.926626,0.89508,0.845148,0.704534,0.414568,0.0]… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/Qwen-Image-2.1-Turbo-GGUF-metrics.NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF-metrics