FinCreditLlama-3.2-3B
๋ชจ๋ธ ๊ฐ์
FinCreditLlama-3.2-3B๋ ๊ธ์ต ์ ์ฉ ํ๊ฐ๋ฅผ ์ํด ํน๋ณํ ์ค๊ณ๋ ํ๊ตญ์ด ์ธ์ด ๋ชจ๋ธ์ ๋๋ค.
๋ฒ ์ด์ค ๋ชจ๋ธ: unsloth/Llama-3.2-3B-Instruct ๋ฐ์ดํฐ์ : himedia/financial_dummy_data_v4 ํ์ต ๋ฐฉ๋ฒ: LoRA (Low-Rank Adaptation) ํ์ต ์ผ์: 20250623_060351
๐ ํ์ต ๊ฒฐ๊ณผ
- Final Training Loss: 0.2510
- Final Validation Loss: 0.2518
- Best Validation Loss: 0.2518 (step 1000)
- Overall Improvement: 89.0%
- Training Time: 47.81 minutes
ํ์ดํผํ๋ผ๋ฏธํฐ
- Learning Rate: 0.0002
- Max Steps: 1000
- Batch Size: 2
- Gradient Accumulation: 8
- LoRA r: 64
- LoRA alpha: 64
- Max Sequence Length: 2048
- Warmup Steps: 5
๐ง ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ๋
- GPU: NVIDIA RTX A5000
- Peak Memory: 4.217 GB
- Memory Usage: 17.9%
์ฌ์ฉ ๋ฐฉ๋ฒ
from transformers import AutoTokenizer, AutoModelForCausalLM
# ๋ชจ๋ธ๊ณผ ํ ํฌ๋์ด์ ๋ก๋
tokenizer = AutoTokenizer.from_pretrained("himedia/fincredit-Llama-3.2-3B-lr2e04-bs16-r64-steps1000-20250623_060351")
model = AutoModelForCausalLM.from_pretrained("himedia/fincredit-Llama-3.2-3B-lr2e04-bs16-r64-steps1000-20250623_060351")
# ๊ฐ๋จํ ์ถ๋ก ์์
prompt = "๊ณ ๊ฐ์ ์ ์ฉ๋ฑ๊ธ์ ํ๊ฐํด์ฃผ์ธ์:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)
๐ ํ์ต ๋ฐ์ดํฐ ํ์ผ
์ด ๋ ํฌ์งํ ๋ฆฌ์๋ ๋ค์ ํ์ต ๊ด๋ จ ํ์ผ๋ค์ด ํฌํจ๋์ด ์์ต๋๋ค:
training_log.json
: ์ ์ฒด ํ์ต ๋ก๊ทธ (JSON ํ์)FinCreditLlama-3.2-3B_20250623_060351_training_curves.png
: ํ์ต ๊ณก์ ์๊ฐํ ์ด๋ฏธ์ง
๋ ํฌ์งํ ๋ฆฌ๋ช ๊ตฌ์ฑ
fincredit-Llama-3.2-3B-lr2e04-bs16-r64-steps1000-20250623_060351 = fincredit-lamma3-4b-lr2e04-bs2-r64-steps1000-20250623_060351
fincredit-lamma3-4b
: ๋ชจ๋ธ ๊ธฐ๋ณธ๋ชlr2e04
: Learning Ratebs2
: Batch Sizer64
: LoRA ranksteps1000
: ํ์ต ์คํ 20250623_060351
: ํ์ต ์๊ฐ
์ฑ๋ฅ
์ด ๋ชจ๋ธ์ ํ๊ตญ์ด ๊ธ์ต ํ ์คํธ์ ๋ํด ํ์ธํ๋๋์ด ์ ์ฉ ํ๊ฐ ๊ด๋ จ ์ง์์๋ต์ ํนํ๋์ด ์์ต๋๋ค.
๋ผ์ด์ ์ค
Apache 2.0
- Downloads last month
- 137
Inference Providers
NEW
This model isn't deployed by any Inference Provider.
๐
Ask for provider support
Model tree for himedia/fincredit-Llama-3.2-3B-lr2e04-bs16-r64-steps1000-new
Base model
meta-llama/Llama-3.2-3B-Instruct
Finetuned
unsloth/Llama-3.2-3B-Instruct