Aether-7B-5Attn-checkpoints

Aether-7B-5Attn — Intermediate Pretraining Checkpoints Aether family** — 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 **Base model:** Aether-7B-5Attn **Instruction-tuned:** Aether-7B-5Attn-it **Live demo (Space):** Aether-Sovereign-AI **Collection:** Aether Foundation Model Part of the Aether Foundation Model collection. License: Apache-2.0. Built by VIDRAFT (주식회사 비드래프트).