prd101-wd/phi1_5-bankingqa-merged

See axolotl config

axolotl version: 0.10.0.dev0

base_model: microsoft/phi-1_5
# optionally might have model_type or tokenizer_type
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

load_in_8bit: false
load_in_4bit: true

datasets:
  - #path: garage-bAInd/Open-Platypus
    path: /workspace/data/alpaca_corrected_bankingqa.jsonl
    type: alpaca

dataset_prepared_path:
val_set_size: 0.1
output_dir: /workspace/outputs/phi-bankingqa-out5


sequence_len: 1024 #reduced to hasten training
sample_packing: true
pad_to_sequence_len: true

#axolotl own suggestion
eval_sample_packing: False

adapter: qlora
#lora_model_dir:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true

wandb_project: phi1.5-bankingqa-finetune
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 8 #increase to hasten training
micro_batch_size: 4
gradient_checkpointing: true #added to hasten training
num_epochs: 1
optimizer: adamw_torch_fused
adam_beta2: 0.95
adam_epsilon: 0.00001
max_grad_norm: 1.0
lr_scheduler: cosine
weight_decay: 0.01 # added to hasten training
learning_rate: 0.0002

bf16: auto
#tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: True
resume_from_checkpoint:
logging_steps: 1
#flash_attention: true
flash_attention: false

warmup_steps: 100
evals_per_epoch: 4
saves_per_epoch: 1
weight_decay: 0.1
resize_token_embeddings_to_32x: true
special_tokens:
  pad_token: "<|endoftext|>"

workspace/outputs/phi-bankingqa-out5

This model is a fine-tuned version of microsoft/phi-1_5 on the /workspace/data/alpaca_corrected_bankingqa.jsonl dataset. It achieves the following results on the evaluation set:

• Loss: 1.1071

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.95) and epsilon=1e-05 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.8635	0.0208	1	1.2920
2.8745	0.2494	12	1.2862
2.7446	0.4987	24	1.2616
2.4361	0.7481	36	1.1899
2.0611	0.9974	48	1.1071

Framework versions

• PEFT 0.15.2
• Transformers 4.51.3
• Pytorch 2.5.1+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1