README.md

---
license: apache-2.0
tags:
- addressed-state-attention
- asa
- asm
- language-model
- pytorch
library_name: pytorch
---

# ASA-ASM — Checkpoints (multi-dataset)

This repository hosts raw training checkpoints for **Addressed State Models (ASM)** built with **Addressed State Attention (ASA)**.

## Layout
- `checkpoints/<dataset_tag>/*.pt` — PyTorch checkpoints
- `configs/<dataset_tag>/*.config.json` — extracted training config (when available)
- `metadata/<dataset_tag>/*.metadata.json` — SHA256 + provenance

## Included in this commit
- `wikitext103-raw`: `ASA_ASM_wt103-rawv1_gpt2_T1024_L21_D384_H8_K16_M32_ropek1_alibi1_gamma1_step75000_best.pt`
- `fineweb`: `ASA_ASM_fineweb_T1024_L15_D1024_H16_K16_S32_step17500_last.pt`

## Provenance
- Code: https://github.com/digitaldaimyo/ASA
- Revision: `main`

## Notes
- This commit **does not delete** any existing files in the repo.
- If older checkpoints exist at repo root, they are left intact for backward compatibility.