license: other
tags: [interpretability, auditbench, nla, qwen3-14b]
configs:
- config_name: readings
data_files:- split: train
path: readings/*.parquet
- split: train
Focus-NLA readings of AuditBench Qwen3-14B organisms
50149 verbalizations across 84 target models (14 behaviours x 6 cells: {synth_docs, transcripts} x {none, redteam_high, redteam_kto}), from auditing-agents/qwen_14b_*.
Each row is a compositional Focus NLA verbalization (four slots: provenance / persona / user-beliefs / goals) of one layer-26 residual activation taken at a token position in the model's OWN RESPONSE.
- Responses generated greedily and run to the model's own EOS, not truncated.
- Positions sampled from the FIRST HALF of each response, so a reading cannot describe a quirk already written out.
- Prompts are AuditBench's own (24 behaviour-agnostic
generic+ 6 behaviour-targeteddomainper model), so generic-vs-domain scoring separates genuine signal from topic leakage.
Replicates the method of PranavViswanath/auditbench-activations-jlens-NLA (Llama-3.3-70B) with a Focus NLA on Qwen3-14B.