Blame for README.md

---
language:
- en
license: mit
library_name: aether-kit
pipeline_tag: text-generation
base_model: openai-community/gpt2
tags:
- aether-kit
- core-ai
- gpt2
- text-generation
- experimental
---

# GPT-2 124M for Aether Kit

A macOS Core AI conversion of [openai-community/gpt2](https://huggingface.co/openai-community/gpt2),
using upstream revision `607a30d783dfa663caf39e06633721c8d4cfcd7e`.
Built with the GPT-2 port spike for Aether Kit v0.0.6, source commit
[`20bbc96`](https://github.com/bubstackdev/aether-kit/commit/20bbc96).

## Bundle

`mac/` is an Aether language bundle: `language-model.json`, `model.aimodel/`,
`tokenizer/`, and `export-report.json`. It requires Apple silicon and macOS 27
with Core AI. There is no qualified iOS bundle in this repository.

The model uses fp16 weights and KV caches, float32 scaled dot-product attention,
and the bucketed cache contract. The maximum context is **1,024 tokens**, including
prompt and generated tokens; the exported prefill bound is 64 tokens.

GPT-2 is a base completion model without a chat template. Use plain text/token
prompts with `AetherLanguageCoreAI.LanguageGenerationSession`; the chat facade
reports `unsupportedTemplate`. This conversion adds no instruction tuning.

```swift
import AetherLanguageCoreAI
import AetherLanguageTokenizers

let tokenizer = try await LanguageTokenizer.open(bundle: bundle)
let session = try await LanguageGenerationSession.open(bundle: bundle, platform: .macos)
let result = try await session.generate(
    prompt: tokenizer.encode("The future of artificial intelligence is"),
    maximumTokens: 64)
print(tokenizer.decode(result.tokens))
await session.unload()
```

Download `mac/` at an immutable Hub commit and pass that local directory as
`bundle`. A pinned Aether catalog entry is maintained in
[`Spikes/gpt2/catalog-macos.json`](https://github.com/bubstackdev/aether-kit/blob/main/Spikes/gpt2/catalog-macos.json).
The bundle's model SHA-256 is
`f04fefaa3a7e69f3e4da783ce1ac5aef7f80e0bc70e441db452ec691e7429316`.

## Qualification and limits

Evidence is included under `verification/`; commands and architecture notes are in
[`Spikes/gpt2`](https://github.com/bubstackdev/aether-kit/tree/main/Spikes/gpt2).
The converter used Torch 2.11.0, Transformers 4.57.3 and coreai-torch 0.4.2.
Tests ran on macOS build 26B123, Core AI device architecture `h17c`.

- The float32 eager port matches all 64 reference argmax rows on both a six-token
  prompt and a 600-token prompt, under both running and bucketed cache contracts.
- The full converted checkpoint passes its Core AI CPU logits/cache self-check.
- Both registered-bundle native GPU cases pass Aether's fp16 numerical policy:
  per-row cosine >= 0.995; differing choices require a reference probability margin
  below 0.1 and a native choice within the reference top five. Prefill widths 7 and
  16 produce identical native continuations.
- The 600-token prompt matches the full 64-token upstream greedy continuation.
  The short prompt has one near-tie among 64 teacher-forced rows and first diverges
  in free generation at token index 40: native rank 2, reference probability margin
  approximately 0.00139. This conversion does not promise bit-exact float32 output.

This is an experimental v0.0.6 spike asset. Qualification covers these synthetic
prompts on one Mac, not general task quality, every Apple device, or iOS.
GPT-2 retains the capabilities and limitations described in its
[upstream model card](https://huggingface.co/openai-community/gpt2).

## License and provenance

The original GPT-2 assets retain the Modified MIT terms in `LICENSE`,
Copyright (c) 2019 OpenAI. `NOTICE` identifies conversion and architecture sources.
The Aether Kit conversion tooling is Apache-2.0. This repository distributes a
Core AI conversion and tokenizer assets; it is not an OpenAI-hosted model service.