mirror of
https://github.com/TheTom/turboquant_plus.git
synced 2026-10-22 11:02:42 -05:00
bee2625d9c
Rewrite to follow Kamradt (2023) and NVIDIA RULER (COLM 2024):
Modes:
- single: Kamradt depth sweep (0-100%) x context length heatmap
- multi-key: RULER MK-NIAH with distractor needles
- multi-value: RULER MV-NIAH with same-key multiple values
Changes from v1:
- "The special magic number is: {7-digit}" format (Kamradt style)
- 24 diverse filler paragraphs (science, history, art, etc.)
- Depth sweep as primary variable (0%, 10%, ..., 100%)
- Heatmap output (the classic red/green grid)
- Distractor needle support (multi-key mode)
- Multi-value retrieval scoring
- 7-digit regex scoring from content field only
- Cites methodology sources in script header
101 tests, all passing. No /Users/tom/ paths.
References:
- github.com/gkamradt/LLMTest_NeedleInAHaystack
- github.com/NVIDIA/RULER (arXiv:2404.06654)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: tturney@psyguard.ai