metadata
base_model:
- meta-llama/Llama-3.1-8B-Instruct
language:
- en
license: mit
library_name: transformers
pipeline_tag: text-generation
Model Card
This is a simulator model used to score candidate natural-language explanations of internal features in Llama-3.1-8B. It was introduced in the paper Training Language Models to Explain Their Own Computations.
Given:
- an input text sequence
x(tokenized), - a candidate explanation
E(e.g., “encodes city names”),
the simulator predicts where the described feature should activate in the sequence (token-level activation scores). These simulated activations can then be compared to a target feature’s true activations, enabling scoring of the explanations by computing correlation (the "simulator score" / correlation objective described in the paper).
- Code: https://github.com/TransluceAI/introspective-interp
- Paper: Training Language Models to Explain Their Own Computations
Usage
Note: This simulator is not usable via standard transformers APIs alone. You must first clone and install the repository, which provides the custom simulator wrapper and scoring utilities.
from observatory_utils.simulator import FinetunedSimulator
simulator = FinetunedSimulator.setup(
model_path="Transluce/features_explain_llama3.1_8b_simulator",
add_special_tokens=True,
gpu_idx=0, # e.g. 0
tokenizer_path="meta-llama/Llama-3.1-8B",
)
Citation
@misc{li2025traininglanguagemodelsexplain,
title={Training Language Models to Explain Their Own Computations},
author={Belinda Z. Li and Zifan Carl Guo and Vincent Huang and Jacob Steinhardt and Jacob Andreas},
year={2025},
eprint={2511.08579},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2511.08579},
}