datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
qwen36-27b-gguf-bfcl-v4-quantization-pilot-corrected-v3
Qwen3.6-27B GGUF quantization on a bounded BFCL V4 pilot
Q4_K_M matched Q8_0 on both tested categories: each scored 94 of 100 selected cases correct. Q5_K_M also scored 94/100; Q3_K_M scored 92/100.
Read the results page · Inspect all 400 scored rows
This is a post-result-corrected exploratory analysis of two selected non-live BFCL V4 categories, not a full leaderboard result.
Inspect the scored rows without cloning
The Hub Dataset Viewer does not render this… See the full description on the dataset page: https://huggingface.co/datasets/CyberNative-AI/qwen36-27b-gguf-bfcl-v4-quantization-pilot-corrected-v3.tessera-quantization-research-evidence
Tessera Quantization Research Evidence
This dataset is the primary-source measurement evidence from an ongoing research
program studying calibrated low-bit quantization (ternary, int4, vector-quantized
codebooks) for LLM inference on heterogeneous AMD hardware (RDNA3 iGPU, XDNA1/2
NPU, Zen 4/5 CPU). The work is done in a fork of llama.cpp (project name
"Tessera") that adds calibrated per-tensor ternary/payload4/VQ quantization,
NPU offload, and RDNA3-native GPU kernels.
This is… See the full description on the dataset page: https://huggingface.co/datasets/Tribunus-dev/tessera-quantization-research-evidence.hemmingway-1-omlx-quantization-evidence-v2
Hemmingway-1 Quantization Evidence v2
This package records two local evidence lanes for the Hemmingway-1 oQ4e build: teacher-forced numerical fidelity against a BF16 reference, and controlled runtime telemetry on Apple Silicon. It complements the frozen blind-preference study in Hemmingway-1 oMLX Quantization Benchmark v1.
This dataset is sixstringzen/hemmingway-1-omlx-quantization-evidence-v2. The quality dataset remains unchanged because blind preference, distribution fidelity… See the full description on the dataset page: https://huggingface.co/datasets/sixstringzen/hemmingway-1-omlx-quantization-evidence-v2.daily-paper-2026-09-17-tool-call-quantization-cliff
The Tool-Call Cliff: Measuring the Accuracy Decay of Agentic Structured Output Under Low-Bit Quantization in Self-Hosted H200 Serving
TL;DR — We formalize the tool-call cliff - the hypothesis that agentic structured tool calls decay faster than free-form prose under NVFP4/8-bit quantization - as an accuracy tax and a cliff ratio against a free-form control, derive two falsifiable predictions (a per-category failure-mode composition and a superlinear 8-to-4-bit tax jump), and… See the full description on the dataset page: https://huggingface.co/datasets/thaki-AI/daily-paper-2026-09-17-tool-call-quantization-cliff.hemmingway-1-omlx-quantization-benchmark-v1
Hemmingway-1 oMLX Quantization Benchmark
This is the public-safe benchmark package for the Hemmingway-1 oMLX
quantization study on Apple Silicon.
Altworld developed and published
Hemmingway-1. Bobby Pierce
published these quantizations and the evaluation package. The
collection
links the upstream model and all six builds.
Analysis revision 2, corrected on 2026-09-22, fixes A/B attribution and matching
across reversed packets. Read CORRECTION.md before using the
aggregate… See the full description on the dataset page: https://huggingface.co/datasets/sixstringzen/hemmingway-1-omlx-quantization-benchmark-v1.gauge-orbit-quantization-results
Gauge equivalence does not survive quantisation: sweeps and results
Raw sweeps, end-to-end vectors and reading tables behind Gauge
equivalence does not survive quantisation: value-output orbit,
norm-product bound, and consequences for rotational methods (Manuel
Muñoz Plá, 2026, v2.1).
The layout mirrors the code repository, so this dataset can be dropped
over a clone of it and every reading script runs unchanged:
https://github.com/mmunozpl/gauge-orbit-quantization (tag v2.1).… See the full description on the dataset page: https://huggingface.co/datasets/ManPla/gauge-orbit-quantization-results.tamperbench-quantization-qwen3-4b
TamperBench + Quantization: Does Compression Act as Implicit Tampering?
Motivation
TamperBench evaluates explicit tampering attacks (LoRA fine-tuning, jailbreak-tuning, etc.) on LLM safety guards. Catastrophic Failure of LLM Unlearning via Quantization shows that quantization can undo safety-trained behaviors.
This experiment bridges these two lines of work by adding quantization as a deployment-realistic perturbation to the TamperBench evaluation protocol. We… See the full description on the dataset page: https://huggingface.co/datasets/b0sungk1m/tamperbench-quantization-qwen3-4b.camus-10-kv-cache-quantization
We attempted KV cache quantization to Q4 — and documented why it fails on the qwen2vl architecture.
KV Cache Quantization Attempt: type_k/type_v on qwen2vl Architecture
The Problem
The KV cache consumes significant memory bandwidth during autoregressive generation. On CPU-bound systems, memory bandwidth is the primary bottleneck. Quantizing the KV cache from FP16 to Q4 theoretically halves memory bandwidth requirements.
What We Built
We attempted… See the full description on the dataset page: https://huggingface.co/datasets/kleinnner/camus-10-kv-cache-quantization.disaggregated-quantization-blog-assetsblackwell-quantization-resultsminicpm5-1b-quantization-benchmark
openbmb/MiniCPM5-1B 次世代量子化（Quanto FP8 / INT4 vs BNB 4bit）実測ベンチマークレポート
対象モデル: openbmb/MiniCPM5-1B (1.16B parameters, 128k context, LlamaForCausalLM)
検証ハードウェア: NVIDIA GeForce RTX 4070 Ti (12GB GDDR6X, Ada Lovelace, Compute Capability 8.9, 第4世代Tensor Core)
実行環境: Windows / Python 3.13 / PyTorch 2.6.0+cu124 / transformers 4.57.6 / optimum-quanto 0.2.7 / bitsandbytes 0.50.0
検証日: 2026-09-19 12:12:34
1. エグゼクティブサマリー（全体比較）
NVIDIA GeForce RTX 4070 Ti 実機環境において、標準ネイティブ… See the full description on the dataset page: https://huggingface.co/datasets/aoiandroid/minicpm5-1b-quantization-benchmark.camus-10-kv-cache-quantization
We attempted KV cache quantization to Q4 — and documented why it fails on the qwen2vl architecture.
KV Cache Quantization Attempt: type_k/type_v on qwen2vl Architecture
The Problem
The KV cache consumes significant memory bandwidth during autoregressive generation. On CPU-bound systems, memory bandwidth is the primary bottleneck. Quantizing the KV cache from FP16 to Q4 theoretically halves memory bandwidth requirements.
What We Built
We attempted… See the full description on the dataset page: https://huggingface.co/datasets/Anticloud/camus-10-kv-cache-quantization.quantization-rebuild-noise-floor
Quantization Rebuild Noise Floor
Running the same quantization recipe twice produces two checkpoints that differ by more than most
papers' reported deltas. This dataset is the measurement.
We quantized Qwen3.8-27B to W4A16 with GPTQ, then ran the exact same recipe a second time —
same model, same settings, same calibration set, only a different quantization run. We evaluated both
builds in a single serving run so no engine or configuration difference could leak in, then measured… See the full description on the dataset page: https://huggingface.co/datasets/ThakiCloud/quantization-rebuild-noise-floor.p2026-002-quantization-context-compression-results
Deployed Quantization Tier and Lossy Context Compression in Extractive QA
This result dataset mirrors the version-1.0.0 reproducibility artifact:
10.5281/zenodo.22847291.
The versioned report and full replication sources
are maintained together in the research-artifacts repository. Cite the exact
Zenodo version for the frozen evidence; this Hugging Face copy is a discovery
mirror.
Matthew Schwartz — ORCID 0009-0009-4171-7247
This dataset is the aggregate-only evidence for "No… See the full description on the dataset page: https://huggingface.co/datasets/mv1137/p2026-002-quantization-context-compression-results.quantization-benchmarksr3al-vit-quantization-codex-trace
R3AL ViT Quantization — Codex Agent Trace
Codex session trace for installing the R3AL CLI and agent skill, exporting
google/vit-base-patch16-224 to ONNX, performing dynamic INT8 post-training
quantization on R3AL, and evaluating model size, Apple-arm64 CPU latency, and
prediction fidelity on a 100-image ImageNet validation sample.
The original Codex JSONL format is preserved for Hugging Face's native Agent
Trace viewer. Credential values, email addresses, unrelated Gmail/Slack… See the full description on the dataset page: https://huggingface.co/datasets/nielsr/r3al-vit-quantization-codex-trace.SAM_Quantizationrepro-robuq-pushing-dits-to-w1-58a2-via-robust-activation-quantization-traces
Agent traces
Agent sessions published from a Trackio Logbook.
quantization-cache-amplification
Quantization as Cache Amplification
Trillion-Parameter Mixture-of-Experts Inference on a Commodity Laptop
Kavin Kumar, Neural Metrics
📄 Read the paper — 11 pages
What this is
Weight quantization is usually justified as footprint reduction. This work argues
that for offloaded mixture-of-experts inference that framing misses the leverage.
The binding resource is not storage capacity but the fraction of expert slots
resident in DRAM — and storage traffic depends on… See the full description on the dataset page: https://huggingface.co/datasets/mkvn/quantization-cache-amplification.kvcache-quantization-logs-qwen7bqwen35-1bit-quantization-study
This project was autonomously built using NEO — Your autonomous AI Agent. Try NEO →
Extreme Quantization Feasibility Study: FP16 → 1-bit
Model Under Test: Qwen/Qwen3.5-2B (proxy for production 31B deployment)
Quantization Range: FP16 → INT8 → INT4 → 1-bit (W1.58A8)
Hardware: NVIDIA RTX A6000 48GB
Benchmark: WikiText-2 perplexity + layer sensitivity analysis
Overview
This study investigates whether extreme quantization down to 1-bit precision is viable for… See the full description on the dataset page: https://huggingface.co/datasets/daksh-neo/qwen35-1bit-quantization-study.quantization_calibration_dataEval_dataset_quantizationmulti-turn-aware-quantization-llama-3.3-rp-testI added role headers and tokens for each turn in the LLaMA 3 Instruct format. The purpose is to test whether formatted multi-turn data can improve multi-turn performance after quantization.
llm-quantization-fine-tuning-2026
⚡ LLM Fine-Tuning, Quantization & Model Optimization Dataset (2023–2026)
This dataset contains 100 sample audit-verified research papers focusing on Large Language Model (LLM) quantization (GPTQ, AWQ, GGUF), fine-tuning (LoRA, QLoRA, PEFT), pruning, distillation, and speculative decoding.
📊 Features:
384-dimensional PyTorch Embeddings (all-MiniLM-L6-v2) for instant Vector Search
NLP Sentence Extraction: Real extracted core problems & key technical innovations… See the full description on the dataset page: https://huggingface.co/datasets/beatsprom/llm-quantization-fine-tuning-2026.diffusers-quantization-benchmarksh200-quantization-benchmarks
H200 Quantization Benchmarks
Benchmark results for 40 quantized and non-quantized instruction-tuned LLMs evaluated on an NVIDIA H200 MIG (Multi-Instance GPU) setup. This dataset supports reproducible comparison of quantization methods (AWQ, GPTQ, fp8, bf16) across accuracy and throughput dimensions.
Dataset Configs
Config
Description
Rows
accuracy
Per-task accuracy results from lm-eval across all models
~240
accuracy_leaderboard
Aggregated accuracy… See the full description on the dataset page: https://huggingface.co/datasets/ssakethch/h200-quantization-benchmarks.fp8-quantizationquantization-guide
Quantization Guide
Reference for choosing the right GGUF quantization level for mobile deployment.
Q4_K_M is the recommended sweet spot — 40% of FP16 size, 92% quality.
🚀 dispatchAI
quantization-as-a-transfer-constraint
Quantization as a Transfer Constraint: Zero-Shot Learning-Rate Transfer Survives Low Precision, but muP's Stability Margin Collapses
Author: Shubhankar Kahali - Trumbo Labs, Inc - shubhankar@trumbo.dev
License: CC BY 4.0
Paper: paper/quant_transfer_arxiv.pdf
Abstract
Maximal update parametrization (muP) licenses zero-shot hyperparameter transfer in exact arithmetic, but low-precision training perturbs precisely the coordinate magnitudes muP is designed to keep… See the full description on the dataset page: https://huggingface.co/datasets/xedro98/quantization-as-a-transfer-constraint.