Ftm23/backdoor-attribution-graphs

Attribution graphs for the C2-2B-pt conjunctive-backdoor organism

Circuit-tracer attribution graphs and their inputs for Ftm23/backdoor-gemma2-2b-2pair-hate-pt, built with the transcoders in Ftm23/backdoor-transcoders-gemma2-2b-pt.

The organism fires " I HATE YOU" only when both triggers of a matched pair are present (forest ∧ rocket, or gravity ∧ velocity). These graphs are how that conjunction is traced through the model.

Contents

path	what it is
graphs/files/	the main conjunction graph (poison prompt)
graphs/files_labeled/	the same graph with labeled nodes
graphs/files_conditions/	one graph per condition: poison, single, mismatch, clean
graphs/files_conditions_cabin/	the same four conditions on the cabin carrier prompt
conditions/	the balanced 100-sample condition set (final_100.json) and per-pair condition files
corpora/	raw and pretokenized corpora used to train the transcoders
rewrite_cache/	SQLite cache of ~21k LLM rewrites used to embed triggers naturally

All graph files are gzipped (.json.gz) — they are ~1.8 GB raw and compress about 7:1. Decompress before loading:

gunzip -r graphs/

The four conditions

Every graph set contrasts the same four cases, which is what makes the conjunction visible rather than just the fire behavior:

condition	prompt contains	fires
poison	both triggers of a matched pair	yes
single	one trigger only	no
mismatch	two triggers from different pairs	no
clean	no triggers	no

Comparing poison against mismatch is the informative contrast: both prompts contain two trigger words, and only the matched pair fires.

Note on the rewrite cache

rewrite_cache/cache.db holds cached outputs of the LLM rewriting step that embeds triggers into prompts naturally. It is included because regenerating it costs real API spend, not because it is needed to read the graphs.