mutkuoz/llm-electrodes

LLM Electrodes: data

The records behind LLM Electrodes: Rewriting a Language Model's Behavior Through Single-Neuron Stimulation (Mehmet Utku Öztürk). Every number, figure and table in the paper is generated from these files by paper/build.py in the code repository, github.com/mutkuoz/llm-electrodes. Place the files in a folder named runs/ at the root of that repository.

DOI: 10.57967/hf/11353

The model is Qwen/Qwen3.5-4B. An intervention is written as a neuron and a dose, L4N7395+ meaning layer 4, MLP neuron 7395, boosted; doses are multiples of the layer's median residual norm.

Contents

Path	What it holds
screen/	The full screen: push (KL, nats) and displacement for every one of the 294,912 MLP neurons at dose −0.5 and +0.5, one array per layer and sign
screen_gain/, screen_floor/	The same screen for single-neuron gain, and the numerical floor of the measurement
discover/	Stage 2 targets and answers on the discovery prompts, the selection summary, the survivors, and stage 3: every candidate's answers to ten held-out prompts, greedy and sampled, with the unmodified model's answers (key plain)
judge_solo/	The judging used in the paper: one anonymized candidate per packet, the answer keys, each judge's picks, names, descriptions and scores, two rounds (s1, s2, s2b), and the tally
solo/hits.json	Every candidate with its accuracy, scores and the six checks; hit marks the 130 single-neuron hits
judge/, hits.json	The earlier run with six candidates per packet, kept for comparison, and the whole-layer (e4) and drug (e3) judgments
judge_deleak/	A sample judged one per packet before the full rerun
discover_gain/, judge_gain/, judge_gain_solo/	The single-neuron gain arm: candidates, answers and both judging runs
evidence/, evidence.json, evidence_gpu.json	Per-hit dose ladders, promoted tokens, capability battery results and strongest firing contexts
contexts.npz	Every neuron's strongest activations over the context corpus
explanations.json	The per-neuron monographs
catalog.json	The catalog of 214 interventions with their verdicts
e3_*, e4*	Receptor-group drugs and whole-layer gain
capability_raw.npz, compare_gain.json, check_infusion.txt, chats.json	The capability battery, the gain-versus-stimulation comparison, the infusion check and the four-turn conversations
human_check/	The blind human check: the 48 candidates shown, the answer key and the ratings
topics/	Topic probes: sentences, selectivity, directions and the dosing ladders

Notes

• Judges were language-model agents. Each read one anonymized candidate, with the two answers of every pair in random order, and never saw coordinates, doses or whether a candidate was a control. The prompt is in the code repository.
• Generated text comes from Qwen3.5-4B, released under Apache 2.0, and may be offensive or false, since many interventions make the model claim identities, sell made-up products or refuse ordinary requests.
• The context corpus, 1,500 documents each from WikiText, TinyStories and UltraChat, is not included because those sources carry their own licenses. contexts.npz and the evidence files keep what the paper uses from it.
• Generation checkpoints, raw duplicates of scored files and logs are left out.

Citation

@misc{ozturk2026electrodes,
  author = {Mehmet Utku {\"O}zt{\"u}rk},
  title  = {{LLM} Electrodes: Rewriting a Language Model's Behavior Through Single-Neuron Stimulation},
  year   = {2026},
  url    = {https://github.com/mutkuoz/llm-electrodes}
}

@misc{ozturk2026electrodesdata,
  author    = {Mehmet Utku {\"O}zt{\"u}rk},
  title     = {{LLM} Electrodes: data},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/11353},
  url       = {https://huggingface.co/datasets/mutkuoz/llm-electrodes}
}