---
tags:
  - coreml
  - apple
  - watchos
  - tinybrain
  - smollm
---

# SmolLM2-135M-Instruct for TinyBrain

This repository packages `HuggingFaceTB/SmolLM2-135M-Instruct` as a compiled Core ML bundle.

## Bundle output

The conversion script produces:

- `smollm2_135m_instruct_only_logits.mlmodelc.zip`

The zip layout is:

- `smollm2_135m_instruct_only_logits.mlmodelc/`
- `tokenizer/`
- `metadata.json`

## Runtime interface

- Task: causal language modeling
- Inputs:
  - `input_ids` (`Int32`, shape `1 x 32`)
  - `attention_mask` (`Int32`, shape `1 x 32`)
- Output:
  - `logits`

## Important

This bundle is marked as `tinybrain-causallm-llama-v1`.

That is intentional: `SmolLM2` uses a Llama-style tokenizer/chat template.

The bundle also includes TinyBrain runtime hints in `metadata.json`:

- `chat_template_type`
- `tokenizer_family`
- `default_system_prompt`
- `sampling_defaults`
- `runtime_profile`

These fields let TinyBrain adapt prompt rendering, history trimming, and sampling defaults without hardcoding the model name in the app.

## Build

Run:

```bash
python3 convert_smollm2_135m_instruct_to_coreml.py
```

The script:

1. Downloads the model and tokenizer from Hugging Face.
2. Converts the model to Core ML.
3. Compiles the package to `.mlmodelc`.
4. Writes runtime metadata.
5. Packages the compiled model, tokenizer, and metadata into one zip.