---
tags:
- artificial-life
- embodied-ai
- mechanistic-interpretability
- babyllm
- pytorch
---

# Project Mochi — published bridge checkpoints

This repository contains **small Project Mochi bridge checkpoints**, not standalone language models.

A Mochling combines a frozen BabyLM ancestor with learned sensorimotor projection layers, action heads, a target-pointer head, episodic-memory projection, caretaker perception, and working interaction state. The bridge checkpoints here are intended to be loaded by the Project Mochi codebase.

## Upstream ancestor

Current published bridges target:

`qyxu1994/instanton-childes-heavy`

The ancestral Transformer remains frozen during Project Mochi bridge training.

## Published artifacts

| Checkpoint | File | SHA-256 | Meaning |
|---|---|---|---|
| MVP-0.7b | `childes_mvp07b_bridge.pt` | `e23ea741d66170298dd16b78b823ceb9006311622d83a22d7364156d1d621235` | Working interaction state + episodic-memory sensorimotor bridge |
| MVP-0.10 | `childes_mvp010_bridge.pt` | `768e476750af3b449ab6739a9801883f29f66532535b6f8ccc9c1c7fc59cb922` | Caretaker embodiment + PUSH + language-conditioned motor control |
| MVP-0.11 | `childes_mvp011_bridge.pt` | `03b754aa1179c046b416e8acf701ab1e7caeda587c2a7877fa09e8b163d63abb` | Environment-dynamics bridge (event novelty, need-modulated reward, seeded layouts) |

Not every experimental run is published. The policy is to publish checkpoints at scientifically meaningful milestone boundaries.

## Usage

The preferred path is through the Project Mochi launcher:

```bash
pip install -e .
mochi-terrarium
```

If the requested named checkpoint is not available locally, the launcher fetches the exact artifact from this repository and records a local SHA-256 provenance sidecar.

Explicit fetch:

```bash
python3 scripts/fetch_checkpoint.py --checkpoint-id mvp-0.10
```

Project source:

`https://github.com/anicka-net/mochlings`

## Scientific scope

### MVP-0.10 language result

MVP-0.10 includes supervised **language-conditioned motor grounding**: matched referential text such as colour words causally shifts the physical target-pointer distribution in held-out synthetic evaluations.

It does **not** yet include a supervised next-token speech objective. The current speech-generation path is an experimental ancestral BabyLM generation conditioned on grounded sensory prefixes and dialogue context. It should not be interpreted as a separately trained grounded speech policy.

### Caretaker embodiment

The caretaker is represented by an explicit structured percept rather than a hidden target override. Pointing, touch, and holding can enter the neural prefix as observable interaction signals.

## Architecture / provenance

The checkpoints contain Project Mochi bridge parameters only. They do **not** vendor the ~27M ancestral model weights.

See the source repository for:

- milestone reports and quantitative evaluation;
- architecture and perception-boundary invariants;
- checkpoint loader implementation;
- exact training curricula;
- third-party licence and dataset-provenance notes.

## Licensing and upstream-data note

The upstream Instanton model code and weights are documented as MIT-licensed (`docs/licensing_third_party.md`). However, the BabyLM training mixture includes upstream corpora with heterogeneous terms, including research/non-commercial restrictions. Do not assume the ancestral training corpus is commercially unencumbered or redistributable.

See `docs/licensing_third_party.md` in the Project Mochi source repository for third-party provenance details.