Skip to content

DATASET_CARD.md

Latest commit

 

History

History
54 lines (35 loc) · 2.67 KB

DATASET_CARD.md

File metadata and controls

54 lines (35 loc) · 2.67 KB
Raw

nanoIM Dataset Card

Summary

nanoIM data is synthetic, symbolic, and generated from source. It is designed to isolate temporal aliasing: paired interactions share the same flattened transcript but require different target actions because of timing, audio state, visual state, policy state, or background/tool state.

Dataset Files

File Purpose
data/mini.jsonl Small smoke suite for quick training and eval.
data/hard.jsonl Deterministic held-out symbolic suite with 10 task families.
data/noisy.jsonl Counterbalanced suite with distractors and timing-only alias pairs.
data/pilot_events.jsonl Scripted local event trace captured with measured wall-clock timing.
data/pilot.jsonl Pilot event trace converted into nanoIM micro-turn examples.

The Hugging Face export additionally writes native split files under data/<suite>/train.jsonl, data/<suite>/validation.jsonl, and data/<suite>/test.jsonl so Hub viewers expose real dataset splits instead of one aggregate JSONL with an internal split column.

Schema

Each JSONL row is one example with:

  • example_id, pair_id, task_family, alias_group_id, template_id
  • flattened_transcript
  • split: train, val, or test
  • critical_timestep
  • turns / timesteps: ordered micro-turns with t_ms, dt_ms, stream fields, and target_action

See docs/schema.md for field-level details.

Task Families

The generator includes yield detection, hesitation handling, barge-in recovery, self-correction, backchannel timing, clarification timing, visual cue trigger, approval interruption, async tool weaving, and background result integration.

Anti-Leakage Boundary

Transcript baselines receive only flattened_transcript. They do not receive IDs, pair IDs, task families, timestamps, event labels, split labels, template IDs, or target labels. Tests and release gates assert that alias-pair transcript features are identical and that transcript paired separation remains 0.00.

Intended Use

  • Study representation loss from turn-based transcripts.
  • Teach and extend micro-turn interaction modeling in a small codebase.
  • Reproduce controlled temporal-aliasing experiments without paid APIs.

Out Of Scope

  • Natural audio or video understanding.
  • Speech recognition, speech synthesis, prosody, or real camera perception.
  • Broad conversational intelligence.
  • Human-subject behavioral claims.

Limitations

The dataset is synthetic by construction. Positive results demonstrate that preserving decisive micro-turn variables matters in this controlled substrate. They do not prove that a tiny model will generalize to unconstrained realtime multimodal interaction.