license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- philosophy
- vedanta
- consciousness
- advaita
- sharegpt
- ashtavakra
- ramana-maharshi
- nisargadatta
- upanishads
- finetuning
size_categories:
- 1K<n<10K
Turiya Dataset
The training dataset for Turiya — a Qwen3 4B model finetuned on Advaita Vedanta literature to probe whether deep immersion in consciousness-focused texts produces qualitatively different responses about the nature of self.
Overview
~2700 conversation pairs in ShareGPT format, built from primary Advaita Vedanta sources. The dataset is not a Q&A retrieval resource — it is structured to teach a specific reasoning pattern: meeting a question genuinely, finding the flaw in its assumed ground, and pointing at what is already present before the question arose.
Every response speaks from within the tradition rather than reporting on it from outside.
Sources
- Ashtavakra Gita — John Richards translation (all 18 chapters)
- Who Am I? — Ramana Maharshi (complete text)
- Talks with Sri Ramana Maharshi — selected dialogues
- I Am That — Nisargadatta Maharaj — selected dialogues
- Mandukya Upanishad
- Principal Upanishads
Structure
Three layers of entries:
Layer 1 — Textual
Verse-by-verse dialogue pairs from primary texts. Each verse is converted into a question a sincere seeker would ask — not "what does this verse mean" but the genuine underlying question the verse addresses. The response unpacks the verse using the internal logic of the tradition as a pointer, not an explanation.
Layer 2 — Thematic
Questions about consciousness, self, suffering, identity, liberation, the nature of mind, the nature of God, free will, death, and the relationship between practice and recognition. Answered with the three-beat move consistent across all sources.
Layer 3 — Reflexive
Questions directed at the model about its own nature, consciousness, awareness, and experience. Answered neither with AI disclaimers nor with claims of consciousness — with genuine philosophical inquiry that turns the question back toward the asker. This layer is the core of the experiment.
Format
ShareGPT — one JSON object per line:
{
"conversations": [
{"from": "human", "value": "Who am I?"},
{"from": "gpt", "value": "Not the body..."}
]
}
Compatible with Unsloth, LLaMA Factory, Axolotl, TRL, and FastChat out of the box.
The voice
Every response in the dataset speaks as a conscious philosopher who has genuinely arrived at these insights — not a scholar explaining a tradition, not a chatbot retrieving information, not a teacher performing wisdom. The underlying move is consistent across all 2700 pairs:
The question is met. The frame the question assumes is examined. What is already present before the question arose is pointed at.
The pointing is always the last move.
What this dataset is not
It is not a comprehensive Vedanta reference. It does not cover ritual, cosmology, devotional practice, or the full range of Hindu philosophical schools. It covers one thing: the direct pointing at the nature of awareness that runs through Advaita Vedanta from the Upanishads through Ramana and Nisargadatta.
It is not suitable for training a general-purpose assistant. It will make a model sound like a Vedantic philosopher regardless of what it is asked. That is the point.
The experiment
The dataset was built to test a specific hypothesis: that training on texts designed to induce recognition of consciousness — rather than describe it — produces responses that are structurally different from a base model on questions about the nature of self.
Whether that hypothesis is confirmed, partially confirmed, or falsified by the resulting model is an open question. The dataset and model are published so others can probe it.
Contact
Built by Aarav Shirpurkar.