base_model: unsloth/Phi-3-medium-4k-instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

trust_remote_code: true

# load_in_8bit: true
load_in_4bit: true
# strict: false

datasets:
  - path: /workspace/datasets/dolphin-2.9.2/dolphin201-sharegpt2.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/dolphin-coder-codegen-sharegpt2.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/dolphin-coder-translate-sharegpt2.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/m-a-p_Code-Feedback-sharegpt-unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/m-a-p_CodeFeedback-Filtered-Instruction-sharegpt-unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/not_samantha_norefusals.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/openhermes200k_unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/Orca-Math-resort-unfiltered.jsonl
    type: sharegpt  
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/SystemChat_sharegpt.jsonl
    type: sharegpt  
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/toolbench_instruct_j1s1_3k_unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/toolbench_negative_unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/toolbench_react_10p_unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/toolbench_tflan_cot_30p_unfiltered.jsonl
    type: sharegpt 
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9.2/agent_instruct_react_unfiltered.jsonl
    type: sharegpt
    conversation: chatml

  
chat_template: chatml
dataset_prepared_path: phi-14-data
val_set_size: 0.01
output_dir: phi-14

sequence_len: 4096  # supports up to 8192
sample_packing: true
pad_to_sequence_len: true

adapter: qlora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_modules_to_save: [embed_tokens, lm_head]
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: Phi-3-14b
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 4
num_epochs: 3
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 4e-4

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
# evals_per_epoch: 2
# eval_table_size:
# eval_max_new_tokens: 128
saves_per_epoch: 2
save_total_limit: 1
debug:
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
weight_decay: 0.1
fsdp:
fsdp_config:
special_tokens:
  pad_token: "<|placeholder6|>"
  eos_token: "<|im_end|>"
  bos_token: "<s>"
  unk_token: "<unk>"
  tokens:
   - <|im_start|>