SeaWolf-AI's picture
sync Aether family cross-links
d04f77f verified
|
Raw
History Blame Contribute Delete
3.07 kB
metadata
license: apache-2.0
language:
  - en
  - ko
tags:
  - aether
  - foundation-model
  - sovereign-ai
  - checkpoints
  - pretraining
  - fully-open
pretty_name: Aether-7B-5Attn Intermediate Pretraining Checkpoints

Aether-7B-5Attn — Intermediate Pretraining Checkpoints

Aether family5Attn Base 5Attn IT 7Attn Base 11Attn Demo Blog Collection

Blog

Intermediate pretraining checkpoints of FINAL-Bench/Aether-7B-5Attn, released for full reproducibility and training-dynamics research — the OLMo-tier "fully open" standard.

Step Tokens (approx) Folder
110,000 ~98B step_110000/
115,000 ~102B step_115000/
162,000 ~144B (final pre-anneal base) step_162000/

The released base model is the annealed 162k checkpoint; these folders are the raw training checkpoints along the way.

Each folder contains model.safetensors + config.json + tokenizer. This is a custom architecture (aether_v2_7way) — load with the aether_pkg/ code shipped in the base model repo (see its "How to load" section). Run inference at batch_size = 1.

Related

Part of the Aether Foundation Model collection. License: Apache-2.0. Built by VIDRAFT (주식회사 비드래프트).