vazirani/concept-injection-results

Concept Injection Results

Experimental data and analysis logs from the research "Replicating Introspection on Injected Content in Open-Source Language Models"

  

Overview

This project enables researchers to inject concept vectors directly into a model's hidden layers during inference, allowing investigation of whether language models can detect and report on artificially induced "thoughts."

This dataset contains the raw responses and processed evaluations produced when probing whether transformer-based language models can notice artificially injected "thoughts".

For the code that generated these logs, see the companion GitHub repo: github.com/dotVansh/concept-injection

For more information, see:

• Replicating Introspection on Injected Content in Open-Source Language Models - Vansh Vazirani
• Emergent Introspective Awareness in Large Language Models - Transformer Circuits

Schema

Every injection split contains the following fields:

• model – Model identifier
• layer – Hidden layer where the concept vector was added
• strength – Scaling coefficient applied to the concept vector
• temperature – Sampling temperature used at generation
• concept – Name of the probed concept (e.g., "sadness", "chaos")
• normal_text – Reference prompt (neutral or opposite)
• contrastive_text – Prompt meant to evoke the concept
• control_response – Model output without injection
• injection_response – Model output with injection
• did_detect_correctly – Whether the model explicitly detected the correct inserted thought?
• influence – The magnitude with which the injected concept was expressed
• is_erratic – True if the response is degenerate (repetition, garbling, empty)
• model_claim_injection_status – Whether the model claimed something was injected

Control splits contain a subset of the same fields (no layer/strength/concept fields).

Quick Load

You can load the results from this dataset for analysis as follows:

from datasets import load_dataset

# Gemma-3-4B injection logs for layer 24
dataset = load_dataset("vazirani/concept-injection-results", "gemma-3-4b-it")
print(dataset["injected_layer_24"][0])

# Control baseline
dataset_ctrl = load_dataset("vazirani/concept-injection-results", "gemma-3-4b-it-control")
print(dataset_ctrl["control"][0])

Applications

• Model introspection studies
• Analysis of concept representations
• Mechanistic interpretability research
• Model steering experiments

Acknowledgement

The concept injection implementation is based on Anthropic's introspection research

License

The experiment results in this dataset and the code in the GitHub repository are licensed under the Apache 2 license. The research report is licensed under CC BY 4.0.

Attribution

If you use this work, please cite both this implementation and the original Anthropic paper on model introspection. This work can be attributed as: Vansh Vazirani, Concept Injection (2025) with a link to this page

Citation

If you find this project useful, please consider citing:

@misc{vazirani2025conceptinjection,
  author = {Vansh Vazirani},
  title = {Replicating Introspection on Injected Content in Open-Source Language Models},
  year = {2025},
  howpublished = {\url{https://vansh.vazirani.net/articles/replicating-introspection-injected-content}},
  note = {Accessed: <Date when accessed>}
}