metrics

Exp 0 - qualitative agreement check (real gemma-2-2b SAE)

Layer 24　·　gemma-2-2b / 24-res-matryoshka-dc　·　blocks.24.hook_resid_post　·　edge: reverse coverage ≥ 0.5, both endpoints fire ≥ 20

For each block pair we compare edges the metrics KEEP (survivors) against edges they REJECT despite passing the crude coverage test. Read the parent/child labels: survivors should be semantically related; rejected edges should look like frequency / co-occurrence artifacts. Labels from Neuronpedia (not fetched - URLs only).

Block pair 0->1

survivor (8)

• 73 -> 446 R=0.94 F=0.13 recon_gain=0.640 recon=Y surv=1.02 p_fires=13%
  • parent 73: technical academic language
  • child 446: words related to events, changes, difficulties, or limitations
• 103 -> 192 R=0.73 F=0.08 recon_gain=0.428 recon=Y surv=1.05 p_fires=4%
  • parent 103: Danish words or phrases, especially those that might be movie titles, as well as parenthetical year dates
  • child 192: code snippets and questions about programming, some of which are in Russian
• 21 -> 449 R=0.71 F=0.02 recon_gain=0.425 recon=Y surv=0.97 p_fires=27%
  • parent 21: programming source code
  • child 449: various syntax elements commonly found in different coding languages like C++, PHP, etc
• 103 -> 493 R=0.91 F=0.15 recon_gain=0.393 recon=Y surv=1.04 p_fires=4%
  • parent 103: Danish words or phrases, especially those that might be movie titles, as well as parenthetical year dates
  • child 493: mixed English and Russian text related to website coding
• 21 -> 465 R=0.78 F=0.00 recon_gain=0.391 recon=Y surv=1.09 p_fires=27%
  • parent 21: programming source code
  • child 465: references to figures and tables
• 124 -> 231 R=0.65 F=0.08 recon_gain=0.326 recon=Y surv=0.90 p_fires=9%
  • parent 124: code with names separated by underscores like “NG X”
  • child 231: capitalized words or acronyms separated by underscores
• 74 -> 491 R=0.81 F=0.16 recon_gain=0.315 recon=Y surv=0.97 p_fires=9%
  • parent 74: technical writing in the field of electrical engineering
  • child 491: text related to patents, inventions, and technical writing
• 33 -> 320 R=0.84 F=0.17 recon_gain=0.314 recon=Y surv=0.98 p_fires=7%
  • parent 33: words or phrases used in a specific, technical or legal context.
  • child 320: noun phrases followed by a number.

reject:superparent (4)

• 32 -> 365 R=1.00 F=0.00 recon_gain=0.573 recon=Y surv=1.00 p_fires=99%
  • parent 32: numbers within a document
  • child 365: citations in research papers
• 32 -> 193 R=1.00 F=0.00 recon_gain=1.102 recon=n surv=1.00 p_fires=99%
  • parent 32: numbers within a document
  • child 193: language common to legal opinions.
• 32 -> 421 R=1.00 F=0.02 recon_gain=1.295 recon=Y surv=1.00 p_fires=99%
  • parent 32: numbers within a document
  • child 421: text related to the game World of Warcraft.
• 32 -> 387 R=1.00 F=0.01 recon_gain=1.145 recon=Y surv=1.00 p_fires=99%
  • parent 32: numbers within a document
  • child 387: contractions with the word “not.”

reject:freq-driven (4)

• 56 -> 460 R=0.82 F=0.23 recon_gain=0.084 recon=Y surv=0.15 p_fires=8%
  • parent 56: numerical data like version numbers or measurements
  • child 460: numbers and citations
• 0 -> 156 R=0.79 F=0.15 recon_gain=0.429 recon=Y surv=0.13 p_fires=2%
  • parent 0: numbers, especially those related to dates and statistics
  • child 156: number sequences
• 12 -> 308 R=0.78 F=0.08 recon_gain=0.041 recon=Y surv=0.20 p_fires=13%
  • parent 12: LaTeX mathematical expressions
  • child 308: hyphens
• 50 -> 186 R=0.75 F=0.07 recon_gain=0.122 recon=n surv=0.39 p_fires=5%
  • parent 50: math calculation problems involving multiplication, remainders, quotients, factors, and divisors
  • child 186: characters and symbols associated with mathematical expressions and code.

reject:no-recon (4)

• 21 -> 173 R=0.85 F=0.18 recon_gain=0.142 recon=n surv=0.99 p_fires=27%
  • parent 21: programming source code
  • child 173: references to legal and scientific publications, including journals, acts, and institutions
• 6 -> 435 R=0.82 F=0.00 recon_gain=0.050 recon=n surv=0.98 p_fires=43%
  • parent 6: words that are part of names of organisms, locations or people and abbreviations commonly used when referring to them.
  • child 435: code snippets and related terms including Russian
• 21 -> 288 R=0.79 F=0.05 recon_gain=0.265 recon=n surv=1.01 p_fires=27%
  • parent 21: programming source code
  • child 288: numbers and date-like sequences, especially those found in log files or code
• 6 -> 211 R=0.78 F=0.04 recon_gain=0.055 recon=n surv=1.00 p_fires=43%
  • parent 6: words that are part of names of organisms, locations or people and abbreviations commonly used when referring to them.
  • child 211: words related to storytelling and imagination in quotes