hprobes 0.4.0

Hallucination neuron discovery and causal validation for transformer LLMs

hprobe

Discover and causally validate hallucination-associated FFN neurons (H-Neurons) in transformer LLMs.

Based on arXiv:2512.01797.

Install

pip install hprobe
# or
uv add hprobe

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer
from hprobe import HProbe

model = AutoModelForCausalLM.from_pretrained("google/gemma-3-4b-it", torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("google/gemma-3-4b-it")

# samples: list of dicts with question, options, answer
probe = HProbe(model, tokenizer)
probe.fit(samples, options_key="choices", answer_key="answer")

print(probe.n_neurons_, probe.layer_distribution_)

results = probe.score()
print(f"AUROC {results['auroc']:.3f}  gap {results['auroc_gap']:+.3f}")

probe.causal_validate()

CLI

# Fit and score on an MCQ dataset
hprobe run --model google/gemma-3-4b-it --data dataset.jsonl --samples 500

# Transfer: score a saved probe on a different model
hprobe transfer --probe results/probe --model google/gemma-3-4b --data dataset.jsonl

# Fit from pre-generated responses with judge labels
hprobe responses --model google/gemma-3-4b-it --data responses.jsonl

Supported formats

Input files: .jsonl, .json, .parquet

Auto-detected dataset formats: mmlu, medqa, medmcqa. Any other format works by passing options_key and answer_key directly.

Key options

Parameter	Default	Description
l1_C	0.01	Inverse L1 strength — lower = fewer neurons
contrastive	True	3-vs-1 labeling at the generated answer token
layer_stride	1	Sample every Nth layer (2 = faster)
validation_split	0.2	Holdout fraction for scoring
max_tokens	1024	Truncation length

Save & load

probe.save("results/gemma_medqa")          # writes .json + .pkl
probe = HProbe.load("results/gemma_medqa", model, tokenizer)
probe.score_on(new_samples, options_key="choices", answer_key="answer")