---
license: other
license_name: mixed-artifact-licenses
license_link: https://huggingface.co/stdlogits/drowse-web-gpt2-base/blob/main/LICENSES.md
base_model: openai-community/gpt2
pipeline_tag: text-generation
library_name: mlc-llm
tags:
- drowse
- webgpu
- interpretability
- base-model
---

# GPT-2 Base for drowse

A raw-completion base model for [drowse.ai](https://drowse.ai), which runs it on the device through WebGPU and reads and steers it
as it generates. This repository holds the converted weights, the tokenizer, the WebAssembly model library with drowse's hooks compiled in, and the core geometry pack (neutral statistics and the default concept manifolds for each context). drowse reads this model at the residual stream; it has no lens or units packs.

- Source: [openai-community/gpt2](https://huggingface.co/openai-community/gpt2/tree/607a30d783dfa663caf39e06633721c8d4cfcd7e)
- Weights: `q0f32`: float32 weights, unquantized
- Context: 1024 tokens
- Runtime: `drowse-web-runtime-v2`, hook ABI `block-sites-v5`, profile `standard-v4-residual`

These are browser artifacts, not a Transformers checkpoint. Install them through [drowse.ai](https://drowse.ai): its signed catalog pins each download to an immutable revision of this repository and verifies every file's SHA-256 digest.

## Reproducibility and licenses

`drowse-build.json` records the exact source files, the MLC LLM base commit and every compiler overlay's
hash; `hosted-artifacts.json` and `hosted-pack-artifacts.json` list every file with its digest. The model
library is compiled by the recipe in `browser-runtime/` of
[stdlogits/drowse.ai](https://github.com/stdlogits/drowse.ai), which pins the upstream TVM, MLC LLM and
WebLLM sources and the patches drowse applies to them. Licenses apply by component: see
[LICENSES.md](LICENSES.md).

[drowse](https://github.com/stdlogits/drowse) is written by a9lim with Claude; its browser edition, [drowse.ai](https://drowse.ai), is built from [stdlogits/drowse.ai](https://github.com/stdlogits/drowse.ai).