metadata

language:
  - sw
license: apache-2.0
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - llama
  - trl
  - sft
base_model: llama-3-8b
datasets: mwitiderrick/SwahiliAlpaca
pipeline_tag: text-generation

Swahili llama 3 8b

Developed by: GodsonNtungi
License: apache-2.0
Finetuned from model : unsloth/llama-3-8b-bnb-4bit

An experimental model with poor performing results, but a great start

training run : 1 epoch
time: 9 hours : 20 mins : 07 seconds
training loss: 0.8683

PEFT parameters

model = FastLanguageModel.get_peft_model(
    model,
    r = 16, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 16,
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
    random_state = 3407,
    use_rslora = False,  # We support rank stabilized LoRA
    loftq_config = None, # And LoftQ
)

Weakness
The model is not properly finetuned to generate end of text token when needed , hence great results start followed by gibberish depending on max token limit set