922-Narra/tagalog-lm-lora-tests

Experimental Tagalog loras: safe or accurate outputs not guaranteed (not for production use)!

Note: better/best results with

• Prompting in Tagalog
• Using format "Human: (prompt)\nAssistant:"

Example: "Ito ay isang chat log sa pagitan ng AI Assistant na nagta-Tagalog at isang Pilipino. Magsimula ng chat:\nHuman: Hello po?\nAssistant:"

lt2_08162023

• Fine tuned on a small dataset of 14 items, manually edited
• 1 epoch (barely any noticable results)
• From chat LLaMA-2-7b
• Lora of chat-tagalog v0.1

lt2_08162023a

• Fine tuned on a small dataset of 14 items, manually edited
• 20 epochs (more observable effects)
• From chat LLaMA-2-7b
• Lora of chat-tagalog v0.1a

lt2_08162023b

• Fine tuned on a small dataset of 14 items, manually edited
• 10 epochs
• From chat LLaMA-2-7b
• Lora of chat-tagalog v0.1b

lt2_08162023c

• Fine tuned on a small dataset of 14 items, manually edited
• 50 epochs (overfitted)
• From chat LLaMA-2-7b
• Lora of chat-tagalog v0.1c

lt2_08162023d

• Fine tuned on a small dataset of 14 items, manually edited
• 30 epochs (v0.1a further trained and cut-off before overfit)
• From chat LLaMA-2-7b
• Lora of chat-tagalog v0.1d

llama-2-7b-tagalog-v0.2 loras (08/26/2023)

• Fine tuned on dataset of ~10k items (mixed)
• 2/2a/2b fine-tuned for 1/2/3 epochs
• From chat LLaMA-2-7b
• Future attempt planned with cleaner chat/dialogue data

hopia-3b-v0.1 (08/26/2023)

• Fine tuned on a small dataset of 14 items, manually edited
• 20 epochs
• From Open LLaMA 3b

llama-2-7b-tagalog-v0.3 loras (09/01/2023)

• Fine tuned on a dataset of ~1k items (Tagalog-focused dataset, based off Tagalog sentences augmented by LLaMA-2-13b base to create a 3-turn dialogue dataset between Human and Assistant)
• 3/3a fine-tuned for 1/2 epochs
• From chat LLaMA-2-7b
• Experiment on partially synthetic data (and observing capability of LLaMA-2 base on generating Tagalog): will be further curating dataset
• Loras for chat-tagalog v0.3) and chat-tagalog v0.3

llama-2-7b-tagalog-v0.3WC2 (09/01/2023)

• Fine tuned on experimental dataset of ~6k items (Tagalog-focused dataset, based off Tagalog sentences and Wiki entries augmented by LLaMA-2-13b to create a dialogue-QnA dataset between Human and Assistant)
• 1 epoch
• From chat LLaMA-2-7b

llama-2-13b-tagalog-v0.3 loras (09/01-02/2023)

• Fine tuned on experimental datasets of ~1k items (Tagalog-focused dataset, based off Tagalog sentences augmented by LLaMA-2-13b base to create a 3-turn dialogue dataset between Human and Assistant)
• 3 fine-tuned for 1 epoch, rank = 16, lora alpha = 32
• 3a with rank = 8
• 3b for 2 epochs
• 3c for 1 epoch, lr = 1e-4, warmup steps = 0.1
• 3d with lr = 2e-4, rank = 32, lora alpha = 64
• 3e for 2 epochs
• From LLaMA-2-13b
• Trying LLaMA-2-13b chat/other base and curated dataset for next attempts