NovoCode/Phi-2-DPO

See axolotl config

axolotl version: 0.4.0

base_model: microsoft/phi-2
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: Intel/orca_dpo_pairs
    type:
      system_prompt: ""
      field_system: system
      field_instruction: question
      field_output: rejected
      field_output: chosen

dataset_prepared_path:
val_set_size: 0.05
output_dir: ./phi-sft-out

sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true

adapter:
lora_model_dir:
lora_r:
lora_alpha:
lora_dropout:
lora_target_linear:
lora_fan_in_fan_out:

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 2
num_epochs: 2
optimizer: adamw_torch
adam_beta2: 0.95
adam_epsilon: 0.00001
max_grad_norm: 1.0
lr_scheduler: cosine
learning_rate: 0.000003

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: True
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

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

phi-sft-out

This model is a fine-tuned version of microsoft/phi-2 on the Intel/orca_dpo_pairs dataset. It achieves the following results on the evaluation set:

• Loss: 1.2999

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: 3e-06
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-05
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
1.3053	0.0	1	1.3288
1.2314	0.25	287	1.3183
1.1664	0.5	574	1.3090
1.4349	0.75	861	1.3034
1.4875	1.0	1148	1.3012
1.3461	1.23	1435	1.3006
1.3247	1.48	1722	1.2998
1.2906	1.73	2009	1.2999

Framework versions

• Transformers 4.37.0
• Pytorch 2.1.2+cu121
• Datasets 2.16.1
• Tokenizers 0.15.0