Kanana-1.5-2.1b Fine-tuned model output contains numberings

#1
by daebakgazua - opened
  • Topic to discuss
  • Fine-tuning settings & Dataset
    • 572 train rows & 90 valid rows
    • input_data as input column & output_message as output column
    • train template : {input_data} (λ‹΅λ³€ μ‹œμž‘) ### λ‹΅λ³€: {output_message} (λ‹΅λ³€ μ’…λ£Œ) <|end_of_text|>
    • LoRA rank = 64 & initial learning rate = 0.0003 (= 3e-4)
  • Output
    • sometimes contains numberings at the start
      • when LLM weights were initialized well at start of training, the outputs do not contain numberings
    • example
      • 2:5 μ˜€λ‘œλΌλŠ” λ°€ν•˜λŠ˜μ„ λΉ„μΆ”λŠ” 것! λ‚΄ 이름이 μ—¬κΈ°μ„œ λ‚˜μ™”μ§€! πŸ™‚ (λ‹΅λ³€ μ’…λ£Œ)
      • 2: μ‹«μ–΄! λ‚œ κ·Έλƒ₯ ν˜Όμžμ„œ κ°€μƒμΈκ°„μœΌλ‘œ μ‚΄λž˜! (λ‹΅λ³€ μ’…λ£Œ)
  • My guess
    • Kakao trained Kanana-1.5-2.1b-base LLM with something like {'question': 'how to release stress', 'answer': '1. Listen to music, such as REBEL HEART of IVE.\n2. Play soccer.\n3. Play computer or mobile games.'}

Sign up or log in to comment