mech-interp
high-temp-refusal-mech-interp
High-Temperature Refusal Mechanistic Interpretability Artifacts
This dataset repository contains the machine-readable artifacts for the gated
Qwen2.5-7B-Instruct mechanistic-interpretability campaign documented in
MECH_INTERP_RESULTS.md.
The core artifacts include:
all-layer greedy residuals and raw logit-lens arrays for 200 paired JBB prompts;
6,600 temperature-conditioned trajectories at seven layers and 17 positions;
all-layer prompt-final residuals for 12,157 deduplicated… See the full description on the dataset page: https://huggingface.co/datasets/amirali1985/high-temp-refusal-mech-interp.mech_interp_experiments
Attack / critic / defender template experiments
Python tooling for prompt-injection template search, evaluation against a target model, and optional defender stress tests.
Start here for humans: skim AGENTS.md (written for AI agents and contributors).
Quick start
Environment: set OPENROUTER_API_KEY and/or provider keys as needed. Optional: .env in this folder, or ../data_creation/.env.
Data: data/merged_dev_with_memories.json is the default dataset for eval, attack, and… See the full description on the dataset page: https://huggingface.co/datasets/psidharth567/mech_interp_experiments.hh_rlhf_with_code_backdoors_combinedsae-llama-the-pile-max-activation-locations-v2llama-mlp8-outputs
llama-mlp8-outputs
Uncompressed .parquet files of unshuffled float-16 floats
(stored as 2 bytes per float!) of the MLP layer activations of llama3.2 1B.
sae-llama-the-pile-max-activation-locations