Upload folder using huggingface_hub

Update canonical Pixel full-corpus perplexity rows

Remove stale duplicate file perplexity/train-00000-of-00001.parquet

Remove stale duplicate file quality_benchmarks/train-00000-of-00001.parquet

Remove stale duplicate file x86_inference/train-00000-of-00001.parquet

Remove stale duplicate file m4_inference/train-00000-of-00001.parquet

Remove stale generated split folder pixel_inference

Update public v1 dataset with validated M4 Qwen and M4 CPU reruns

Update README: add Qwen x86, fix row counts (4407 total, m4=1021, x86=399)

Update pixel_inference.parquet: integrate Qwen x86 + M4 CPU cliff data (2026-04-11)

Update m4_inference.parquet: integrate Qwen x86 + M4 CPU cliff data (2026-04-11)

Update x86_inference.parquet: integrate Qwen x86 + M4 CPU cliff data (2026-04-11)

fix(DR2): M4 CPU TPS sweep ordering; Q5_K_M cliff confirmed cliff-prone

fix(DR2): add TPS sweep rows (context_len=0, n=10), 1,019->1,026 rows

data: Q5_K_M cliff rerun n=5 (2026-04-10), 2,875->2,930 rows

feat: add BoolQ + TruthfulQA imatrix rows for Q2_K and Q3_K_M (103→105 rows)

feat: add x86 cliff sweep data (385 rows, ctx=256-2048, n=5 trials per variant)

feat: add M4 CPU cliff sweep rows (ngl=0, 4 threads, 88 rows, 2026-04-09)

feat: add M4 CPU cliff sweep data (2026-04-09) — 1,019 rows, 4,398 total

audit: Q3_K_M cliff-attenuated (not immune); Q4_K_M/Q5_K_M thermal artifact caveat

audit: correct speedup ~99%, imatrix deltas, totals 4310, quality_benchmarks 105 rows

docs: update README — correct row counts (x86: 7→392, quality: 103→107, total: 3923→4312), fix Key Findings (Q3_K_M cliff-immune, imatrix hurts low-bitwidth, x86 AVX2 replication), update x86 methodology to describe n=5 cliff sweep

Upload dataset

Upload dataset

Upload dataset

Upload dataset

Upload dataset

Upload folder using huggingface_hub

initial commit