quantbench-artifacts — the exact quantized weights behind the QuantBench rows

These are the actual artifacts produced by the 2026-07-30/31 QuantBench sweep, salvaged off the Modal volume so leaderboard rows can be verified against real weights instead of taken on trust.

Canonical set: each (model x method) at calibration wikitext2-train, n=128, seed 0. The other 68 artifacts from the sweep were left on the volume; these four cover the headline comparisons.

Directory	Base model	Method	Calibration	Base licence
qwen25-1p5b__gptq__wikitext2__n128__s0	Qwen/Qwen2.5-1.5B-Instruct	GPTQ	wikitext2-train n=128 seed 0	Apache-2.0
qwen25-1p5b__awq__wikitext2__n128__s0	Qwen/Qwen2.5-1.5B-Instruct	AWQ	wikitext2-train n=128 seed 0	Apache-2.0
smollm2-1p7b__gptq__wikitext2__n128__s0	HuggingFaceTB/SmolLM2-1.7B-Instruct	GPTQ	wikitext2-train n=128 seed 0	Apache-2.0
smollm2-1p7b__awq__wikitext2__n128__s0	HuggingFaceTB/SmolLM2-1.7B-Instruct	AWQ	wikitext2-train n=128 seed 0	Apache-2.0

Provenance and honesty notes

• Quantization ran on an NVIDIA A10; evaluation ran on A10 and T4.
• GPTQ: g128 desc_act=True damp=0.01 calib_len=2048, gptqmodel-7.3.2, torch 2.8.0+cu128 / transformers 5.14.1.
• AWQ: zero_point g128 GEMM calib_len=512, autoawq-0.2.9, torch 2.6.0+cu124 / transformers 4.51.3 (autoawq is deprecated upstream and was pinned to its last-tested combination).
• Calibration windows were drawn with a seeded RNG from wikitext-2-raw-v1 train. Seed 0 here; seeds 1 and 2 exist in the measured rows.
• Every GPTQ row in this sweep loaded via torch-fallback, not an optimized kernel — GPTQ throughput numbers in the leaderboard measure a dequantize-in-PyTorch path and are not a kernel benchmark.
• These weights are derivatives of Apache-2.0 base models; the base licence and attribution carry over. Nothing here was human-reviewed for quality.

Measurements: Mohaaxa/quantbench-sweep (rows.csv + per-row JSON + logs).