RDson/Phi-3-mini-code-finetune-128k-instruct-v1

Finetune of microsoft/Phi-3-mini-128k-instruct on m-a-p/CodeFeedback-Filtered-Instruction for ~9-10h using a single 3090 24GB.

Due to limited resources and time, the training was only on half (0.5136) of the epoch.

  train_loss: 0.43311

    learning_rate=5e-5,
    lr_scheduler_type="cosine",
    max_length=1024,
    max_prompt_length=512,
    overwrite_output_dir=True,
    beta=0.1,
    gradient_accumulation_steps=8,
    optim="adamw_torch",
    num_train_epochs=1,
    evaluation_strategy="steps",
    eval_steps=0.2,
    logging_steps=1,
    warmup_steps=50,
    fp16=True,
    save_steps=50