---
license: mit
library_name: glucofm
tags:
  - cgm
  - glucose
  - foundation-model
  - jepa
  - onnx
  - safetensors
  - gguf
  - rlxp
  - healthcare
pipeline_tag: feature-extraction
---

# GlucoFM encoder weights

Frozen **encoder-only** checkpoints for
[GlucoFM](https://arxiv.org/abs/2605.30865) (dual-stream CGM foundation model),
converted for the Rust [`glucofm`](https://crates.io/crates/glucofm) / RLX stack.

> **Not medical advice.** Research reimplementation. Official Google GlucoFM
> weights are not public. This package redistributes / converts a community
> OpenCGM reconstruction.

## Files

| File | Format | Notes |
|------|--------|-------|
| `glucofm_encoder.safetensors` | F32 safetensors | **Preferred** for `glucofm` load / train warm-start |
| `glucofm_encoder.f16.safetensors` | F16 safetensors | ~2× smaller disk |
| `glucofm_encoder.rlxp` | RLX package F32 | Flat mmap weight-only pack (`rlxp` / `infer` feature) |
| `glucofm_encoder.f16.rlxp` | RLX package F16 | ~2× smaller RLXP |
| `glucofm_encoder.f32.gguf` | GGUF F32 | Lossless GGUF dump |
| `glucofm_encoder.f16.gguf` | GGUF F16 | Smaller disk; dequantized at load |
| `glucofm_encoder.q8.gguf` | GGUF Q8_0 | ~4× smaller than F32; dequantized at load |
| `glucofm_encoder.bin` | Native `GLFM` v1 | Compact glucofm-native dump |
| `glucofm_encoder.onnx` | ONNX opset 17 | Upstream OpenCGM export (raw grid → 128-d embedding) |
| `glucofm_encoder.onnx.meta.json` | JSON | Upstream export metadata |
| `config.json` | JSON | Geometry + param counts for this package |
| `SHA256SUMS` | text | Integrity checks |

**Encoder params in weight files:** 435,505 (OpenCGM ONNX tensors).  
`online/mask_token` (128) is omitted upstream; `glucofm` inference defaults it to zeros.  
Full encoder including `mask_token` + `online/gaussian.rho`: **435,633**.

Paper full trainable target (encoder + JEPA heads): **732,593**.

## Geometry

| Symbol | Value |
|--------|------:|
| Grid `L` | 288 (5 min) |
| Patches `P` | 24 |
| Patch len `K` | 12 |
| Hidden `D` | 128 |
| Stream `S` | 64 |

## Provenance

- Upstream community encoder: [`sfourdrinier/opencgm-stateevent`](https://huggingface.co/sfourdrinier/opencgm-stateevent)
- OpenCGM project: [sfourdrinier/opencgm](https://github.com/sfourdrinier/opencgm)
- Paper: [arXiv:2605.30865](https://arxiv.org/abs/2605.30865)
- Conversion via `glucofm` `convert_weights --all` (ONNX → safetensors / GGUF / RLXP / `.bin`)

Upstream ONNX meta (epoch 40 seed 17, raw stats, circadian, learnable σ) is kept in
`glucofm_encoder.onnx.meta.json`.

## Use with `glucofm`

```bash
cargo run --release --features infer --example infer -- \
  --weights /path/to/glucofm_encoder.safetensors

# RLXP (same feature set)
cargo run --release --features infer --example infer -- \
  --weights /path/to/glucofm_encoder.rlxp

cargo run --release --features hub --example train_from_hub -- \
  --weights /path/to/glucofm_encoder.safetensors --steps 50
```

```rust,ignore
use glucofm::prelude::*;

let mut model = GlucoFm::for_inference(GlucoFmConfig::default(), InferOptions::default(), &mut rng)?;
model.load_weights_path("glucofm_encoder.rlxp")?; // or .safetensors / .gguf / .bin
let emb = model.encode_embedding_day(&day, &mut rng)?; // [128]
```

## Formats

- Prefer **safetensors** or **RLXP F32** for numeric fidelity with RLX `Param` graphs.
- **GGUF** / **RLXP F16** are for disk savings; runtime still uses dense F32/F16 params after load
  (no in-graph Q8 matmul).

## License

Weights and packaging in this repository: **MIT** (see `LICENSE`).

Upstream OpenCGM / paper licenses may impose additional terms on the original
checkpoints — review those projects before commercial use.