Instructions to use FINAL-Bench/Aether-7B-5Attn-it with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use FINAL-Bench/Aether-7B-5Attn-it with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="FINAL-Bench/Aether-7B-5Attn-it", trust_remote_code=True) messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# Load model directly from transformers import AutoModelForCausalLM model = AutoModelForCausalLM.from_pretrained("FINAL-Bench/Aether-7B-5Attn-it", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use FINAL-Bench/Aether-7B-5Attn-it with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "FINAL-Bench/Aether-7B-5Attn-it" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "FINAL-Bench/Aether-7B-5Attn-it", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/FINAL-Bench/Aether-7B-5Attn-it
- SGLang
How to use FINAL-Bench/Aether-7B-5Attn-it with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "FINAL-Bench/Aether-7B-5Attn-it" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "FINAL-Bench/Aether-7B-5Attn-it", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "FINAL-Bench/Aether-7B-5Attn-it" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "FINAL-Bench/Aether-7B-5Attn-it", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use FINAL-Bench/Aether-7B-5Attn-it with Docker Model Runner:
docker model run hf.co/FINAL-Bench/Aether-7B-5Attn-it
license: apache-2.0
language:
- ko
- en
base_model: FINAL-Bench/Aether-7B-5Attn
pipeline_tag: text-generation
library_name: transformers
tags:
- aether
- mixture-of-experts
- korean
- vidraft
- sft
Aether-7B-5Attn-it
๐ Aether Foundation Model ์ปฌ๋ ์ ์ ์ผ๋ถ โ ๋ฒ ์ด์คยทinstructยท์ฒดํฌํฌ์ธํธ๋ฅผ ํ๊ณณ์์.
๐งฉ ์ค๊ฐ ์ฒดํฌํฌ์ธํธ (110k ยท 115k ยท 162k)๋ ๋ฐ์ดํฐ์ ์ผ๋ก ๊ณต๊ฐ: Aether-7B-5Attn-checkpoints. Aether-7B-5Attn base ๋ฅผ ์ง๋ํ์ต(SFT)ํ ํ์ฒ๋ฆฌ ๋ฒ์ ์ ๋๋ค. ๊ฐ๊ด์/๋ฒค์น๋งํฌ ์๋ต ํ์์ ๋ง์ถฐ ํ๋๋์์ต๋๋ค. 6.59B MoE (ํ์ฑ ์ฝ 3B), 49์ธต, 7ร7 ๋ผํด ๋ฐฉ์ง ์ดํ ์ ๋ฐฐ์น. Apache-2.0.
ํ์ต๋ฅ (LR) ์ ํ โ loss๊ฐ ์๋๋ผ held-out ์ ๋ต๋ฅ ๋ก
full-parameter SFT๋ฅผ ์ธ ํ์ต๋ฅ (2e-6 / 6e-6 / 2e-5)๋ก ๊ฐ๊ฐ ์ํํ ๋ค, ํ์ต loss๊ฐ ์๋๋ผ held-out ๋ฒค์น๋งํฌ ์ ๋ต๋ฅ ๋ก ์ต์ข ๋ณธ์ ์ ํํ์ต๋๋ค. (์์ ๋ชจ๋ธ์์ ๋์ LR์ ํ์ต loss๋ ๋ฎ์ถ์ง๋ง ์ค์ ๋ฅ๋ ฅ์ ์์์ํฌ ์ ์์ด, loss๋ก ๊ณ ๋ฅด๋ฉด ์ ๋ฉ๋๋ค.)
| LR | K-AI 4 ํ๊ท | ๋น๊ณ |
|---|---|---|
| 2e-6 | 29.7% | ๊ณผ์์ ํฉ (์ผ๋ถ ๊ณผ๋ชฉ ์ฝํจ) |
| 6e-6 (์ ํ) | 34.9% | ์ต๊ณ ยท๊ท ํ, ์์ ์์ |
| 2e-5 | 32.3% | ํ ๊ณผ๋ชฉ ๊ธ๋ฝ (ํ์ ์์ ์งํ) |
Held-out ๊ฒฐ๊ณผ (๋ณธ ์ ํ๋ณธ, 6e-6)
- GPQA-Diamond (198): 25.3%
- K-AI 4 ํ๊ท (195): 34.9%
- musr_ko 26.5% ยท com2_main_ko 50.0% ยท click 32.0% ยท kommlu_pro 30.4%
- base(SFT ์ ) ๋๋น: base 26.7% โ 34.9% (+8.2pp) โ ๋์ผ held-outยท๋์ผ ํ๋์ค๋ก ์ธก์
๋ชจ๋ ํ๊ฐ๋ ํ์ต์ ์ฐ์ด์ง ์์ held-out ์ ์ ๋๋ค. SFT๋ MMLU-auxiliary(๊ฐ๊ด์ ํ์) ์์์ ์ํ๋์๊ณ , ํ๊ฐ ๊ณผ๋ชฉ(GPQAยทK-AI)๊ณผ ๊ฒน์น์ง ์์ต๋๋ค.
์ฌ์ ํ์ต ๋ฐ์ดํฐ ๊ตฌ์ฑ (๋ฒ ์ด์ค ์์, ์ ํจ ์ํ๋ง ๋น์ค)
| ๋๋ฉ์ธ | ๋น์ค |
|---|---|
| Math (finemath + open-web-math) | 37.8% |
| Korean (webtext + synth) | 21.6% |
| English webยทsynthetic (fineweb-edu + cosmopedia) | 21.6% |
| Code (opc) | 13.5% |
| phase15 ํ๋ฆฌ๋ธ๋ ๋ | 5.4% |
์ด ๊ตฌ์ฑ์ ๋ฒ ์ด์ค ์ฌ์ ํ์ต ๋ฐ์ดํฐ์ ๋น์ค์ ๋๋ค. ๋ณธ ๋ชจ๋ธ์ ํ์ฒ๋ฆฌ(SFT) ๋ฐ์ดํฐ๋ MMLU-auxiliary(๊ฐ๊ด์)์ ๋๋ค.
์ํคํ ์ฒ (๋ฒ ์ด์ค ๊ณ์น)
๋ณธ ๋ชจ๋ธ์ Aether-7B-5Attn base ์ ๋์ผํ ์ํคํ ์ฒ์ ๋๋ค: 49์ธต์ 7ร7 ๋ผํด ๋ฐฉ์ง์ผ๋ก ๋ฐฐ์นํ ์ด์ข ์ดํ ์ (7๋ผ๋ฒจ / 5๋ฉ์ปค๋์ฆ) + MoE(25์ ๋ฌธ๊ฐ top-7). ๊ตฌ์กฐ ์์ธ๋ ๋ฒ ์ด์ค ๋ชจ๋ธ ์นด๋ ยง3.2 ์ฐธ์กฐ.
์ฌ์ฉ
โ ๏ธ batch_size=1 ๋ก ์ฌ์ฉํ์ธ์. ๋ณธ ๋ชจ๋ธ์ ์ผ๋ถ ์ดํ ์ (NSA ๊ณ์ด)์ padding mask ๋ฅผ ์ฌ์ฉํ์ง ์์ผ๋ฏ๋ก, ์ฌ๋ฌ ์ํ์ค๋ฅผ padding ํ์ฌ ๋ฐฐ์น๋ก ๋ฃ์ผ๋ฉด ๊ฒฐ๊ณผ๊ฐ ์กฐ์ฉํ ์ค์ผ๋ ์ ์์ต๋๋ค. ํ ๋ฒ์ ํ๋์ฉ(batch_size=1) ์ถ๋ก ํ์ญ์์ค.
๋ฒ ์ด์ค์ ๋์ผํ๊ฒ aether_pkg/ ์ ์ปค์คํ
์ํคํ
์ฒ๋ฅผ ์ฌ์ฉํฉ๋๋ค. ๋ก๋ฉ ๋ฐฉ๋ฒ์
FINAL-Bench/Aether-7B-5Attn ์ ๋ฐ๋ฅด์ญ์์ค.
์คํ์์ค ์์ ๊ณต๊ฐ LLM ๋น๊ต (6๊ฐ๊ตญ)
๋ฒ ์ด์ค Aether ๊ธฐ์ค. ๋น๋๋ํ๋ ์ ์ผํ ํ๊ตญ AI ์คํํธ์ ์ด๋ฉฐ 5์ข (์ต๋ค) ์ดํ ์ ยท๋ผํด๋ฐฉ์ง ๊ตฌ์กฐ๊ฐ ์ฐจ๋ณ์ ์ ๋๋ค.
๋ฌธ์
์ฃผ์ํ์ฌ ๋น๋๋ํํธ (VIDRAFT) ยท arxivgpt@gmail.com

