File size: 2,074 Bytes
950061c 685f8cc 950061c 685f8cc 950061c 685f8cc 950061c 685f8cc 950061c 685f8cc 950061c 685f8cc 950061c 685f8cc |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 |
---
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](https://huggingface.co/papers/2511.08579).
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](https://github.com/TransluceAI/introspective-interp)
- **Paper:** [Training Language Models to Explain Their Own Computations](https://huggingface.co/papers/2511.08579)
---
## Usage
**Note:** This simulator is not usable via standard `transformers` APIs alone. You must first **clone and install [the repository](https://github.com/TransluceAI/introspective-interp/tree/main#)**, which provides the custom simulator wrapper and scoring utilities.
```python
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
```bibtex
@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},
}
``` |