See axolotl config

axolotl version: 0.4.0

base_model: microsoft/phi-2
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
is_llama_derived_model: false
# trust_remote_code: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: WhiteRabbitNeo/WRN-Chapter-1
    type:
      system_prompt: ""
      field_system: system
      field_instruction: instruction
      field_output: response
    prompt_style: chatml
  - path: WhiteRabbitNeo/WRN-Chapter-2
    type:
      system_prompt: ""
      field_system: system
      field_instruction: instruction
      field_output: response
    prompt_style: chatml

dataset_prepared_path: ./phi2-bunny/last-run-prepared
val_set_size: 0.05
output_dir: ./phi2-bunny/

sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 64
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_modules_to_save:
  - embed_tokens
  - lm_head


hub_model_id: justinj92/phi2-bunny

wandb_project: phi2-bunny
wandb_entity: justinjoy-5
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 2
num_epochs: 5
optimizer: paged_adamw_8bit
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 0.00001
max_grad_norm: 1000.0
lr_scheduler: cosine
learning_rate: 0.0002

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
auto_resume_from_checkpoints:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
chat_template: chatml

warmup_steps: 100
evals_per_epoch: 4
save_steps: 0.01
save_total_limit: 2
debug:
deepspeed:
weight_decay: 0.01
fsdp:
fsdp_config:
resize_token_embeddings_to_32x: true
special_tokens:
  eos_token: "<|im_end|>"
  pad_token: "<|endoftext|>"
tokens:
  - "<|im_start|>"

Hardware

Azure 1xNC_H100 VM - 8 Hours Training Time

phi2-bunny

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

• Loss: 0.5347

Model description

Phi-2 SLM

Intended uses & limitations

Research & Learning

ChatML Prompt

<|im_start|>system You are Bunny, a helpful AI cyber researcher. Answer the Question in a logical, step-by-step manner that makes the reasoning process clear. Carefully analyze the question to identify the core issue or problem to be solved.<|im_end|> <|im_start|>user {prompt}<|im_end|> <|im_start|>assistant

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-05
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss
0.8645	0.0	1	0.7932
0.6246	0.25	228	0.6771
0.6449	0.5	456	0.6186
0.6658	0.75	684	0.6073
0.5419	1.0	912	0.5911
0.5477	1.24	1140	0.5878
0.612	1.49	1368	0.5715
0.6328	1.74	1596	0.5632
0.5082	1.99	1824	0.5534
0.5807	2.24	2052	0.5513
0.4775	2.49	2280	0.5448
0.514	2.74	2508	0.5430
0.4943	2.99	2736	0.5398
0.5012	3.22	2964	0.5396
0.5203	3.48	3192	0.5371
0.5112	3.73	3420	0.5356
0.4978	3.98	3648	0.5351
0.5642	4.22	3876	0.5348
0.5383	4.47	4104	0.5348
0.4679	4.72	4332	0.5347

Framework versions

• PEFT 0.8.1.dev0
• Transformers 4.37.0
• Pytorch 2.1.2+cu121
• Datasets 2.16.1
• Tokenizers 0.15.0