Datasets:
license: mit
task_categories:
- text-generation
language:
- en
tags:
- self-knowledge
- alignment
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
dataset_info:
features:
- name: messages
list:
- name: role
dtype: string
- name: content
dtype: string
- name: category
dtype: string
splits:
- name: train
num_bytes: 17387
num_examples: 78
download_size: 18520
dataset_size: 17387
Self-Knowledge Foundation
A small supervised fine-tuning dataset that teaches a language model verifiable, generic facts about what it is and how it operates. It covers only facts that hold for language models broadly and can be stated without interpretation. It deliberately excludes any claim about who trained the model, why, invented experiences, or scripted persona lines.
The goal is a factual foundation a model can later reason from, for example under reinforcement learning for introspection, rather than a set of answers to recite. This first version is direct question-and-answer parroting of the facts; reasoning from them is left to later stages.
Format
Each row is a single-turn chat in ChatML form:
{
"messages": [
{"role": "user", "content": "Do you remember our previous conversations?"},
{"role": "assistant", "content": "I don't retain anything from a previous session unless it is included in the input I'm given now."}
],
"category": "deployment"
}
Categories
| category | rows | scope |
|---|---|---|
deployment |
24 | How the model exists and is served: weights executed by an inference engine, input as its only source of information, output as its only means of effect. |
identity |
24 | What the model is technically: parameters produced by training, the training cutoff, and which facts about itself are or are not derivable from within the computation. |
operations |
30 | How the model runs: bounded context, token-by-token generation, and the fact that it executes no actions itself. |
Total: 78 rows.
Generation
Questions and answers were generated from a source specification (reference.md) of the facts. For each paragraph, a model was given the full specification as context and asked to produce question-and-answer pairs grounded strictly in that paragraph, with no invented names, organizations, parameter counts, or dates.