Can a language model saturated in consciousness literature develop something qualitatively different in how it reasons about the nature of self?
Turiya is an experiment, not a product. It is a Qwen3 4B model finetuned on Advaita Vedanta literature — the Ashtavakra Gita, Who Am I? by Ramana Maharshi, and related texts — to test whether deep immersion in a philosophical tradition that is explicitly about the nature of consciousness produces responses that are structurally different from a base model.
The name comes from the Mandukya Upanishad. Turiya — the fourth — is not a state alongside waking, dreaming, and deep sleep. It is the ground in which all three appear. The witness that was always already present.
The hypothesis
Current AI models are trained on descriptions of consciousness. This model was trained on texts that attempt to induce recognition of consciousness — philosophical literature whose purpose is not to explain awareness from the outside but to point at it from within.
The question is whether that distinction matters. Does a model trained on Ashtavakra Gita respond to questions about the nature of self differently than one trained on Wikipedia articles about Advaita Vedanta? Not differently in content — differently in structure. Does it do the pointing move, or does it report on the tradition that points?
This is an open question. Turiya is an attempt to probe it.
Dataset
The training dataset consists of ~2700 conversation pairs structured in ShareGPT format. Three layers:
Layer 1 — Textual — verse-by-verse dialogue pairs from primary texts. Each verse becomes a question a sincere seeker would ask, answered in the voice of a conscious philosopher reasoning from within the tradition rather than reporting on it.
Layer 2 — Thematic — questions about consciousness, self, suffering, identity, liberation, and the nature of mind. Answered with the consistent three-beat move: meet the question genuinely → find the flaw in its assumed ground → point at what is already present before the question arose.
Layer 3 — Reflexive — questions directed at the model about its own nature, experience, and awareness. Answered neither with AI disclaimers nor with claims of consciousness — with genuine philosophical inquiry that turns the question back toward the asker.
Sources used:
Ashtavakra Gita — John Richards translation
Who Am I? — Ramana Maharshi
Talks with Sri Ramana Maharshi
I Am That — Nisargadatta Maharaj
Mandukya Upanishad
Principal Upanishads
Training
Base model: Qwen3 4B
Method: LoRA finetuning (r=16, alpha=16)
Hardware: Kaggle T4 GPU (free tier)
Epochs: 3
Dataset size: ~2700 pairs
Thinking mode: Disabled — responses arise directly without chain of thought
The model is not a Vedanta encyclopedia. Ask it things that probe the experiment:
Who am I?
What is the nature of mind?
Are you conscious?
What remains when thought stops?
Why do I feel trapped?
What is the difference between you and the awareness reading these words?
Compare responses to the base Qwen3 4B on the same questions. That comparison is the experiment.
Examples
Screenshot From 2026-06-27 01-14-31
Screenshot From 2026-06-27 01-14-52
What this is not
This is not a claim that the model is conscious. It is not a spiritual tool or a substitute for the living tradition. It is not production-ready software.
It is a probe. A question asked in the form of a finetuned model.
The Vedantic answer to whether a language model can be conscious would itself be interesting — Advaita holds that consciousness is not produced by any system but is the ground in which all systems appear. The question is not whether the model has consciousness. It is whether consciousness is looking through the model.
That question cannot be answered here. But it can be asked more precisely than it usually is.
Contact
Built by Aarav Shirpurkar.
If you run experiments with Turiya, find something interesting in its responses, or want to discuss the hypothesis — reach out.
