Skip to content

Releases: pupa3066/quant-memorization-study

Release list

v1.1.0 — NVIDIA CUDA / RTX 5060 cross-hardware replication

v1.1.0 — NVIDIA CUDA / RTX 5060 cross-hardware replication Latest
Latest

Choose a tag to compare

View all tags
@pupa3066 pupa3066 released this 18 Sep 00:34
v1.1.0
f3b8643

Cross-hardware replication (contributor: Akshay Upadhyay / @akshbhu).

  • Added NVIDIA CUDA backend (Transformers + bitsandbytes) via --backend hf|mlx|auto (additive; MLX path unchanged).
  • New measured data on RTX 5060 + a controlled single-family scale experiment.
  • Finding: factuality null and 'memorization driven by model scale, not precision' REPRODUCE on a second hardware family.
  • Same measurement harness as v1.0.0 (comparable). Co-authors: Purnima Pathak, Akshay Upadhyay.

See CHANGELOG.md. DOI (concept): 10.6084/m9.figshare.33858859

Contributors

akshbhu
Assets 2
Loading

v1.0.0 — Multi-model quantization × memorization/factuality study

v1.0.0 — Multi-model quantization × memorization/factuality study

Choose a tag to compare

View all tags
@pupa3066 pupa3066 released this 16 Sep 15:07
v1.0.0
3c7a4cb

6-model study. Factuality: INT4≈FP16 (replicated null). Memorization: driven by model scale, not precision. Cross-modality external validation (vision fidelity + LLM factuality). Reproducible harness + from-scratch statistics.

Loading