kimhyeongjun/Hermes-3-Llama-3.1-8B-Korean-Finance-Advisor
This is my personal toy project for Chuseok(Korean Thanksgiving Day).
This model is a fine-tuned version of NousResearch/Hermes-3-Llama-3.1-8B on the Korean_synthetic_financial_dataset_21K.
Model description
Everything happened automatically without any user intervention.
Based on finance PDF data collected directly from the web, we refined the raw data using the 'neuralmagic/Meta-Llama-3.1-70B-Instruct-FP8' model. After generating synthetic data based on the cleaned data, we further evaluated the quality of the generated data using the 'meta-llama/Llama-Guard-3-8B' and 'RLHFlow/ArmoRM-Llama3-8B-v0.1' models. We then used 'Alibaba-NLP/gte-large-en-v1.5' to extract embeddings and applied Faiss to perform Jaccard distance-based nearest neighbor analysis to construct the final dataset of 21k, which is diverse and sophisticated.
λͺ¨λ κ³Όμ μ μ¬μ©μμ κ°μ μμ΄ μλμΌλ‘ μ§νλμμ΅λλ€.
μΉμμ μ§μ μμ§ν κΈμ΅ κ΄λ ¨ PDF λ°μ΄ν°λ₯Ό κΈ°λ°μΌλ‘, λμ΄ μμ΄μ 'neuralmagic/Meta-Llama-3.1-70B-Instruct-FP8' λͺ¨λΈμ νμ©νμ¬ Raw λ°μ΄ν°λ₯Ό μ μ νμμ΅λλ€. μ μ λ λ°μ΄ν°λ₯Ό λ°νμΌλ‘ ν©μ± λ°μ΄ν°λ₯Ό μμ±ν ν, 'meta-llama/Llama-Guard-3-8B' λ° 'RLHFlow/ArmoRM-Llama3-8B-v0.1' λͺ¨λΈμ ν΅ν΄ μμ±λ λ°μ΄ν°μ νμ§μ μ¬μΈ΅μ μΌλ‘ νκ°νμμ΅λλ€. μ΄μ΄μ 'Alibaba-NLP/gte-large-en-v1.5'λ₯Ό μ¬μ©νμ¬ μλ² λ©μ μΆμΆνκ³ , Faissλ₯Ό μ μ©νμ¬ μμΉ΄λ 거리 κΈ°λ°μ κ·Όμ μ΄μ λΆμμ μνν¨μΌλ‘μ¨ λ€μνκ³ μ κ΅ν μ΅μ’ λ°μ΄ν°μ 21kμ μ§μ ꡬμ±νμμ΅λλ€.
Task duration
3days (20240914~20240916)
evaluation
Nothing (I had to take the holiday off.)
sample
Framework versions
- Transformers 4.44.2
- Pytorch 2.4.0+cu121
- Datasets 2.21.0
- Tokenizers 0.19.1
- Downloads last month
- 48