jlov7/nanoim-microturn-tiny

nanoIM MicroTurn Tiny Models

This model repository package contains tiny from-scratch PyTorch checkpoints for the nanoIM symbolic temporal-aliasing lab.

What Is Included

Path	Description
checkpoints/tiny/best.pt	Mini-suite GRU checkpoint.
checkpoints/hard/best.pt	Hard-suite GRU checkpoint.
checkpoints/noisy/best.pt	Noisy-suite GRU checkpoint.
checkpoints/transformer_hard/best.pt	Hard-suite tiny Transformer checkpoint.
configs/*.yaml	Training configs used to produce the checkpoints.
reports/*.json	Scorecards, sweeps, controls, and release verification evidence.

Parameter Counts

Checkpoint	Parameters
checkpoints/tiny/best.pt	19,288
checkpoints/hard/best.pt	65,604
checkpoints/noisy/best.pt	115,256
checkpoints/transformer_hard/best.pt	257,988

Largest model is ~258K parameters. All others are under 120K.

Intended Use

Use these checkpoints to reproduce nanoIM scorecards and inspect how a native micro-turn representation separates alias pairs that a transcript-only baseline cannot separate.

The models consume symbolic micro-turn features. They are not text generators, not chat models, and not production assistants.

Loading

The nanoIM evaluator loads checkpoints with torch.load(..., weights_only=True).

From the source repository, evaluate the committed source-tree checkpoint:

uv run python -m nanoim.eval --checkpoint runs/noisy/full/seed_7/best.pt --suite noisy --data data/noisy.jsonl --out reports/noisy_scorecard.json

From this generated Hugging Face model repository, install or clone the nanoIM source package, then point the evaluator at the model-repo checkpoint path:

uv run python -m nanoim.eval --checkpoint checkpoints/noisy/best.pt --suite noisy --data ../dataset/data/noisy.jsonl --out reports/noisy_scorecard.json

The generated model repo also includes MANIFEST.json and SHA256SUMS so a reviewer can bind checkpoint files to the release manifest before evaluation.

Evaluation Summary

Model	Suite	TAA	Delta vs transcript oracle
MicroTurn Tiny GRU	hard	1.00	+0.50
MicroTurn Tiny GRU	noisy	1.00	+0.50
MicroTurn Tiny Transformer	hard	1.00	+0.50

The rule harness and a memorized field-lookup table also reach 1.00. The result is about the transcript representation's 0.50 ceiling, not about model superiority.

Transcript-only paired separation remains 0.00 on the hard and noisy suites.

Limitations

The checkpoints prove a controlled symbolic representation result. They do not perform ASR, TTS, visual perception, natural language generation, tool execution, or realtime dialogue management.