jlov7/nanoim-symbolic

nanoIM Symbolic Temporal Aliasing Dataset

nanoIM is a small, synthetic, symbolic dataset for studying temporal aliasing in interaction models. Paired examples can share the same flattened transcript while requiring different target actions because timing, overlap, visual cues, policy events, or background/tool results differ.

Files

File	Hub config	Purpose
data/mini/{train,validation,test}.jsonl	mini	Quick smoke suite for training and evaluation.
data/hard/{train,validation,test}.jsonl	hard	Deterministic ten-family symbolic suite.
data/noisy/{train,validation,test}.jsonl	noisy	Counterbalanced distractor/timing suite used for ablations and controls.
data/pilot/test.jsonl	pilot	Scripted local event trace converted into two alias pairs.

The generated Hub bundle also includes source aggregate files such as data/noisy.jsonl for exact source-repo parity. Hub dataset configs use the split files above so train, validation, and test are native dataset splits.

data/pilot_events.jsonl is included as source evidence for the pilot importer, but the pilot dataset config loads data/pilot.jsonl.

Schema

Each JSONL row contains:

• identifiers: example_id, pair_id, alias_group_id, template_id;
• task metadata: task_family, split, critical_timestep;
• transcript field: flattened_transcript;
• micro-turn fields: turns and timesteps, including t_ms, dt_ms, stream fields, and target_action.

The transcript-only boundary intentionally excludes pair IDs, task families, timestamps, event labels, split labels, template IDs, and target labels. Anti-leakage tests in the source repository enforce this boundary.

Intended Use

• Evaluate whether a model needs native micro-turn features to separate balanced alias pairs.
• Teach temporal aliasing with a small, inspectable dataset.
• Reproduce the paper-style experiments in the nanoIM source repo.

Out Of Scope

• Natural speech, audio, or video understanding.
• Production dialogue quality.
• Human-subject behavioral claims.
• Any claim that the pilot trace is a natural realtime benchmark.

Citation

Use the citation metadata from the source repository CITATION.cff when citing nanoIM.