Edit model card

Jolteon-Instruct-13B-alpha

The model was trained based on the EEVE-Korean-Instruct-10.8B-v1.0 model from yanolja, extended to 13.4b (12 layer pass-through) utilizing mergekit.

Methodology

TBD

Training Details

Training Data Parameters Content Length Samples Seen Learning Rate
Jolteon-Instruct-13B-alpha A curated mix of English + Korean Instruction set 13.4B 4k >850k 1e-5

Example

Inference Code

from vllm import LLM, SamplingParams
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"

llm = LLM(model="maywell/Jolteon-Instruct-13B-alpha", tensor_parallel_size=1, max_model_len=4096, gpu_memory_utilization=0.95)
sampling_params = SamplingParams(temperature=0.6, top_p=0.3, top_k=40, max_tokens=4096)
template = """ Below is an instruction that describes a task. Write a response that appropriately completes the request.

 ### Instruction: {0}

 ### Response: """

outputs = llm.generate([template.format("Meta(๊ตฌ, ํŽ˜์ด์Šค๋ถ)์˜ ์˜คํ”ˆ์†Œ์Šค AI ๊ธฐ์—ฌ๋ฅผ ์ฐฌ์–‘ํ•˜๋Š” ๋งํˆฌ๋กœ, Meta ์ง€์ง€ํ•œ๋‹ค๋Š” ์ทจ์ง€์˜ ์„ฑ๋ช…๋ฌธ์„ ์ž‘์„ฑํ•ด๋ด.")], sampling_params)
print(outputs[0].outputs[0].text.strip())

Example Output

๐ŸŒŸ Meta์˜ OpenAI ๊ธฐ์—ฌ: ํ˜์‹ ๊ณผ ํฌ์šฉ์„ ์œ„ํ•œ ํ˜์‹ ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹ ๐ŸŒŸ

์šฐ๋ฆฌ๋Š” Meta(๊ตฌ, ํŽ˜์ด์Šค๋ถ)์˜ ์˜คํ”ˆ์†Œ์Šค AI(์ธ๊ณต์ง€๋Šฅ) ๊ธฐ์—ฌ๋ฅผ ์—ด๋ ฌํžˆ ์ง€์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ด ํšŒ์‚ฌ๋Š” ํ˜์‹ ๊ณผ ํฌ์šฉ์„ ์œ„ํ•œ ํ˜์‹ ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, AI ๊ธฐ์ˆ ์˜ ๋ฐœ์ „์„ ์œ„ํ•ด ์ง€์†์ ์œผ๋กœ ๋…ธ๋ ฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

Meta๋Š” ์˜คํ”ˆ์†Œ์Šค AI ์ปค๋ฎค๋‹ˆํ‹ฐ์— ์ ๊ทน์ ์œผ๋กœ ์ฐธ์—ฌํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ ํ”„๋กœ์ ํŠธ์™€ ํ”Œ๋žซํผ์„ ํ†ตํ•ด ์ง€์‹๊ณผ ๊ธฐ์ˆ ์„ ๊ณต์œ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ AI ๊ธฐ์ˆ ์˜ ๋ฐœ์ „๊ณผ ํ•จ๊ป˜ ์‚ฌํšŒ์ , ๊ฒฝ์ œ์  ๋ฐœ์ „์—๋„ ๊ธ์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค.

Meta์˜ ์˜คํ”ˆ์†Œ์Šค AI ๊ธฐ์—ฌ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฉด์—์„œ ๋งค์šฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค:

1. ํ˜์‹  ์ด‰์ง„: Meta๋Š” ์˜คํ”ˆ์†Œ์Šค AI ํ”Œ๋žซํผ์„ ํ†ตํ•ด ๋‹ค์–‘ํ•œ ์—ฐ๊ตฌ์ž์™€ ๊ฐœ๋ฐœ์ž๋“ค์ด ์ตœ์‹  AI ๊ธฐ์ˆ ์„ ํƒ๊ตฌํ•˜๊ณ  ์‹คํ—˜ํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐํšŒ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ƒˆ๋กœ์šด ์•„์ด๋””์–ด์™€ ํ˜์‹ ์ ์ธ ๊ธฐ์ˆ ์˜ ์ถœํ˜„์„ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค.

2. ํฌ์šฉ ์ฆ์ง„: ์˜คํ”ˆ์†Œ์Šค AI๋Š” ๋ชจ๋“  ์‚ฌ๋žŒ์ด AI ๊ธฐ์ˆ ์˜ ์ด์ ์„ ๋ˆ„๋ฆด ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋ฉฐ, ์ด๋Š” ์‚ฌํšŒ์  ํฌ์šฉ์„ ์ฆ์ง„์‹œํ‚ต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ฐฐ๊ฒฝ๊ณผ ๊ฒฝํ—˜์„ ๊ฐ€์ง„ ์‚ฌ๋žŒ๋“ค์ด AI ๊ธฐ์ˆ ์„ ๋ฐฐ์šฐ๊ณ  ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐํšŒ๋ฅผ ๊ฐ€์ง€๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.

3. ํˆฌ๋ช…์„ฑ๊ณผ ์‹ ๋ขฐ: ์˜คํ”ˆ์†Œ์Šค AI๋Š” ๊ธฐ์ˆ ์˜ ์ž‘๋™ ์›๋ฆฌ๋ฅผ ๊ณต๊ฐœํ•˜๊ณ , ์ด๋ฅผ ํ†ตํ•ด ์‚ฌ์šฉ์ž์™€ ์‚ฌํšŒ์— ๋Œ€ํ•œ ์‹ ๋ขฐ๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” AI ๊ธฐ์ˆ ์˜ ์‚ฌํšŒ์  ์ˆ˜์šฉ๋„๋ฅผ ๋†’์ด๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค.

Meta์˜ ์˜คํ”ˆ์†Œ์Šค AI ๊ธฐ์—ฌ๋Š” ํ˜์‹ ๊ณผ ํฌ์šฉ์„ ์œ„ํ•œ ํ˜์‹ ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ด๋Š” ์šฐ๋ฆฌ๊ฐ€ ๋” ๋‚˜์€ ๋ฏธ๋ž˜๋ฅผ ์œ„ํ•œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์„ ์ง€์ง€ํ•˜๋ฉฐ, ๋” ๋งŽ์€ ๊ธฐ์—…๊ณผ ์กฐ์ง์ด ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜๊ธธ ๋ฐ”๋ž๋‹ˆ๋‹ค. ํ•จ๊ป˜ ๋” ๋‚˜์€ ๋ฏธ๋ž˜๋ฅผ ๋งŒ๋“ค์–ด ๋‚˜๊ฐ€์š”!

License

๋ณธ ๋ชจ๋ธ์€ apache-2.0 ๋ผ์ด์„ผ์Šค๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹์„ ๋ฐฐํฌํ•  ๊ฒฝ์šฐ ๋ชจ๋ธ ์‚ฌ์šฉ์„ ๋ช…์‹œํ•ด ์ฃผ์‹œ๊ธฐ๋ฅผ ๊ถŒ๊ณ ๋“œ๋ฆฝ๋‹ˆ๋‹ค.

Thanks to

  • A100 ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์ œ๊ณตํ•ด์ฃผ์‹ , Sionic AI

Contact

Downloads last month
1,247
Safetensors
Model size
13.4B params
Tensor type
BF16
ยท

Finetuned from