purplesquirrelnetworks/thot-pocket-gaze

Thot Pocket Gaze Behavior Dataset

Training data for Thot Pocket -- an AI avatar eye contact intelligence system that generates naturalistic gaze behavior during conversations.

Main repo: github.com/ExpertVagabond/thot-pocket Crate: crates.io/crates/thot-pocket (v0.1.0)

The training pipeline is included in the GitHub repo under train/ — a GazeTransformer model (4-layer, 128-dim, 3 output heads).

Purpose

Human eye contact follows complex, culturally-informed patterns during conversation. Speakers and listeners avert gaze at different rates, in different directions, and for different durations depending on cognitive load, conversational role, and cultural norms. This dataset captures timestamped gaze annotations from real and simulated conversations to train AI avatars that exhibit believable eye contact behavior rather than the uncanny-valley stare of most virtual agents.

Schema

Each row in the JSONL data files contains:

Field	Type	Description
timestamp_ms	int	Milliseconds from conversation start
conversation_id	string	Unique conversation identifier
speaker_role	string	Who is producing the current speech segment: avatar or user
conversation_state	string	Current avatar state: idle, listening, thinking, speaking
user_gaze_zone	string	Where the user is looking: at_me, away, down
avatar_looking_at_user	bool	Whether the avatar is making eye contact
aversion_direction	string | null	If avatar is not looking at user, the aversion direction: up_right, up_left, right, left, down_right, down_left. Null when avatar_looking_at_user is true.
contact_duration_ms	int	How long the current gaze state has been held (ms)
culture	string	Cultural context label: western, east_asian, middle_eastern, latin_american, south_asian, african
transcript_segment	string	The speech content at this timestamp

Conversation States

• idle -- No active conversation; avatar is in resting gaze pattern
• listening -- User is speaking; avatar should maintain higher eye contact
• thinking -- Avatar is formulating a response; gaze aversion increases naturally
• speaking -- Avatar is delivering speech; gaze follows natural speaker patterns

Gaze Zone Labels

User gaze zones observed from the avatar's perspective:

• at_me -- User is looking at the avatar's face/eyes
• away -- User is looking to the side or at something else
• down -- User is looking downward (phone, notes, etc.)

Aversion Directions

When the avatar breaks eye contact, the direction of gaze shift:

• up_right -- Associated with visual construction/imagination
• up_left -- Associated with visual recall
• right -- Associated with auditory construction
• left -- Associated with auditory recall
• down_right -- Associated with kinesthetic processing
• down_left -- Associated with internal dialogue

Data Files

data/
  train.jsonl    # Training split

How to Load

from datasets import load_dataset

dataset = load_dataset("purplesquirrelnetworks/thot-pocket-gaze")

Contributing Data

We welcome contributions of annotated gaze behavior data. To contribute:

1. Fork this repository
2. Add your annotated data in JSONL format following the schema above
3. Ensure each row includes all required fields
4. Include a culture label for cultural context
5. Open a pull request with a description of your data source and annotation methodology

Annotation Guidelines

• Timestamps should be relative to conversation start (first utterance = 0ms)
• Gaze zone labels should be annotated at a minimum of 100ms granularity
• Aversion direction must be null when avatar_looking_at_user is true
• Cultural labels should reflect the cultural background of the human participant
• Transcript segments should capture the speech content at the annotation timestamp

Data Sources We Accept

• Webcam-recorded conversations with manual gaze annotation
• Eye-tracker data from conversation studies
• Expert-annotated synthetic conversation scenarios
• Published research datasets re-formatted to this schema (with appropriate licensing)

License

MIT

Citation

@dataset{thot_pocket_gaze_2026,
  title={Thot Pocket Gaze Behavior Dataset},
  author={Purple Squirrel Networks},
  year={2026},
  url={https://huggingface.co/datasets/purplesquirrelnetworks/thot-pocket-gaze},
  license={MIT}
}