---
license: bsd-3-clause
library_name: mlx
pipeline_tag: text-generation
base_model: hugohrban/progen2-base
tags:
- progen
- progen2
- protein-language-model
- mlx
- mlx-lm
- safetensors
- bfloat16
- icl-many-replication
widget:
- text: 1MEVVIVTGMSGAGK
- text: 1MKTLLILAV
---

# ProGen2 Base BF16 MLX Conversion

This repository contains a BF16 MLX-LM conversion of
[`hugohrban/progen2-base`](https://huggingface.co/hugohrban/progen2-base), a
mirror of the base ProGen2 model by Nijkamp et al.

The source checkpoint was downloaded from Hugging Face revision
`71228cbfca5960f9fab5775f378bba3673af9f00` and converted from FP32 to BF16
safetensors. The converted tensor file contains 764,803,616 BF16 parameters.

This conversion was prepared for the standalone replication artifact for
*Many Next Token Predictors are In-context Learners*.

## Files

- `model.safetensors`: BF16 converted ProGen2-base weights.
- `progen2_mlx.py`: MLX-LM custom model implementation.
- `config.json`, tokenizer files, and ProGen custom-code files copied from the
  source model with the MLX model-file entry added.

## Loading with MLX-LM

```python
from mlx_lm import load

model, tokenizer = load("N8Programs/ProGen2-base-bf16")
```

## Original Model

Please cite and follow the terms of the upstream ProGen2 work and the source
model repository:

- Source model: https://huggingface.co/hugohrban/progen2-base
- Paper: https://arxiv.org/abs/2206.13517