Small LLM1

์žฌ๋ฏธ๋กœ ๋งŒ๋“  ์•ฝ 5M(4.98M) ํŒŒ๋ผ๋ฏธํ„ฐ์งœ๋ฆฌ ์ดˆ์†Œํ˜• ์–ธ์–ด ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.

โš ๏ธ ์žฌ๋ฏธ๋กœ ๋งŒ๋“  ๋ชจ๋ธ์ด๋ผ ์„ฑ๋Šฅ์€ ๋‚ฎ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ ์šฉ๋„๋กœ ์“ฐ๊ธฐ์—๋Š” ๋ถ€์กฑํ•˜๊ณ , ์งง์€ ์˜์–ด ๋™ํ™” ๋ฌธ์žฅ์„ ์ด์–ด ์“ฐ๋Š” ์ •๋„๋งŒ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

โš ๏ธ ํŒŒ์ธํŠœ๋‹(instruction tuning / chat tuning)์€ ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์‚ฌ์ „ํ•™์Šต๋งŒ ํ•œ base ๋ชจ๋ธ์ด๋ผ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ฑฐ๋‚˜ ๋Œ€ํ™”ํ•˜์ง€ ๋ชปํ•˜๊ณ , ์ฃผ์–ด์ง„ ๋ฌธ์žฅ์„ ์ด์–ด์„œ ์“ฐ๊ธฐ๋งŒ ํ•ฉ๋‹ˆ๋‹ค.

๋ชจ๋ธ ์ •๋ณด

ํ•ญ๋ชฉ ๊ฐ’
๊ตฌ์กฐ Llama ์Šคํƒ€์ผ decoder-only Transformer (์ฒ˜์Œ๋ถ€ํ„ฐ ํ•™์Šต)
ํŒŒ๋ผ๋ฏธํ„ฐ 4.98M
๋ ˆ์ด์–ด / hidden / heads 6 / 256 / 8
FFN ํฌ๊ธฐ 512
์–ดํœ˜ ํฌ๊ธฐ 4,096 (์ง์ ‘ ํ•™์Šตํ•œ byte-level BPE)
์ปจํ…์ŠคํŠธ ๊ธธ์ด 256 ํ† ํฐ
ํ•™์Šต ๋ฐ์ดํ„ฐ TinyStories ์•ž๋ถ€๋ถ„ ์•ฝ 40๋งŒ ๊ฐœ ์ด์•ผ๊ธฐ (์•ฝ 92M ํ† ํฐ)
ํ•™์Šต 6,000 step ร— 64 ร— 256 ํ† ํฐ, RTX 5070 1์žฅ์—์„œ ์•ฝ 4๋ถ„
๊ฒ€์ฆ loss 1.77 (perplexity ์•ฝ 5.9, TinyStories ๊ฒ€์ฆ ๊ตฌ๊ฐ„)

์‚ฌ์šฉ๋ฒ•

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("aaaaaaaieuyd/Small-LLM1")
tok = AutoTokenizer.from_pretrained("aaaaaaaieuyd/Small-LLM1")

x = tok("Once upon a time", return_tensors="pt")
out = model.generate(**x, max_new_tokens=80, do_sample=True, temperature=0.8, top_k=40)
print(tok.decode(out[0], skip_special_tokens=True))

์ถœ๋ ฅ ์˜ˆ์‹œ:

Once upon a time, there was a little girl named Lily. She had a little tooth that was very important to her. The tooth was her special tooth, and she could always count on it to show it to her friends.

ํ•œ๊ณ„

  • ์˜์–ด ๋™ํ™”์ฒด ๋ฌธ์žฅ๋งŒ ํ•™์Šตํ•ด์„œ ํ•œ๊ตญ์–ด๋‚˜ ๋‹ค๋ฅธ ์ฃผ์ œ๋Š” ๊ฑฐ์˜ ๋ชป ๋‹ค๋ฃน๋‹ˆ๋‹ค.
  • ์‚ฌ์‹ค ์ •๋ณด, ์ถ”๋ก , ์ฝ”๋“œ ๊ฐ™์€ ๊ฑด ๊ธฐ๋Œ€ํ•˜๋ฉด ์•ˆ ๋ฉ๋‹ˆ๋‹ค.
  • ์•ž๋’ค๊ฐ€ ์•ˆ ๋งž๋Š” ์ด์•ผ๊ธฐ๋ฅผ ์ž์ฃผ ๋งŒ๋“ญ๋‹ˆ๋‹ค.
  • ํŒŒ์ธํŠœ๋‹์„ ํ•˜์ง€ ์•Š์•„์„œ ์ง€์‹œ๋ฅผ ๋”ฐ๋ฅด์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค.

English: Small LLM1 is a ~5M-parameter toy language model trained from scratch on TinyStories, just for fun. Performance is low, and it is a base model with no fine-tuning (no instruction or chat tuning). It can only continue short, simple English stories.

Downloads last month
143
Safetensors
Model size
4.98M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Dataset used to train aaaaaaaieuyd/Small-LLM1