---
language:
- en
license: mit
library_name: pixel
pipeline_tag: text-generation
tags:
- pixel
- causal-lm
- local-llm
- pytorch
---

# pixel_100m

This repository contains a PIXEL checkpoint exported for use with the PIXEL codebase. It includes the model checkpoint, tokenizer files, exported config metadata, and this model card.

## What This Model Is

`pixel_100m` is a decoder-only transformer checkpoint from the PIXEL project. This bundle is intended to be used with the PIXEL runtime rather than the Transformers `AutoModel` API.

## Architecture

- Approximate parameter class: `~76,466,688`
- Vocab size: `1262`
- Context length: `1024`
- Layers: `12`
- Hidden size: `768`
- Attention heads: `12`
- Key/value heads: `4`
- Intermediate size: `2048`
- RoPE base: `500000`
- Uses MoE: `False`

## Included Files

- `latest.pt`: PIXEL checkpoint
- `manifest.json`: exported checkpoint pointer
- `pixel_tokenizer.model`: SentencePiece tokenizer model
- `pixel_tokenizer.vocab`: SentencePiece tokenizer vocab
- `pixel_model_config.json`: exported typed model config
- `pixel_training_config.json`: exported training config when available

## Training Snapshot

- Training size preset: `100m`
- Total steps saved in checkpoint: `10`
- Sequence length: `32`
- Batch size: `1`
- Gradient accumulation: `2`

## Runtime Snapshot

- Device: `cpu`
- GPU count: `0`
- Dtype: `torch.float32`

## Usage With PIXEL

Clone the PIXEL codebase, place or download this bundle, then run:

```bash
python infer.py --model checkpoints/pixel_100m/latest.pt --prompt "Hello from PIXEL"
```

Make sure the checkpoint and tokenizer come from the same export bundle.

## Limitations

- This checkpoint is not guaranteed to be instruction-tuned.
- Output quality depends on the training corpus and training duration used for this run.
- This bundle is PIXEL-specific and is not advertised as a drop-in Transformers checkpoint.

## Export Provenance

- Source checkpoint: `latest.pt`
- Source tokenizer model: `pixel_tokenizer.model`
- Source tokenizer vocab: `pixel_tokenizer.vocab`