## model
base_model: /home/dgxuser/workspace/Mango/models/Vulpecula
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

## qlora
load_in_8bit: false
load_in_4bit: false
strict: false
## Lora
adapter: lora
lora_model_dir:
lora_r: 64
lora_alpha: 32
lora_dropout: 0.0
peft_use_rslora: true
lora_target_modules:
lora_mlp_kernel: true
lora_qkv_kernel: true
lora_o_kernel: true
lora_target_linear: true
#lora_target_modules: 
#  - gate_proj
#  - down_proj
#  - up_proj
#  - q_proj
#  - v_proj
#  - k_proj
#  - o_proj
## data 
datasets:
  - path: PocketDoc/Dans-Personamaxx-VN
    type: dan-chat-advanced-llama3
  - path: NewEden/LIMARP-Complexity
    type: dan-chat-advanced-llama3
  - path: NewEden/PIPPA-Mega-Filtered
    type: dan-chat-advanced-llama3
  - path: NewEden/OpenCAI-ShareGPT
    type: dan-chat-advanced-llama3
  - path: NewEden/Creative_Writing-Complexity
    type: dan-chat-advanced-llama3
  - path: NewEden/Light-Novels-Roleplay-Logs-Books-Oh-My-duplicate-turns-removed
    type: dan-chat-advanced-llama3
  - path: PocketDoc/Dans-Failuremaxx-Adventure-3
    type: dan-chat-advanced-llama3
  - path: NewEden/Books-V2-ShareGPT
    type: dan-chat-advanced-llama3
  - path: NewEden/Deepseek-V3-RP-Filtered
    type: dan-chat-advanced-llama3
  - path: NewEden/Final-Alpindale-LNs-ShareGPT
    type: dan-chat-advanced-llama3
  - path: NewEden/DeepseekRP-Filtered
    type: dan-chat-advanced-llama3
  - path: NewEden/RP-logs-V2-Experimental
    type: dan-chat-advanced-llama3
  - path: anthracite-org/kalo_opus_misc_240827
    type: dan-chat-advanced-llama3
  - path: anthracite-org/kalo_misc_part2
    type: dan-chat-advanced-llama3
  - path: NewEden/Storium-Prefixed-Clean
    type: dan-chat-advanced-llama3
shuffle_merged_datasets: true
dataset_prepared_path: base-dataset_prepared
val_set_size: 0.0
output_dir: ./SFT-Vulpecula

## Liger + CCE
plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: false
cut_cross_entropy: true

## CTX settings
sequence_len:  8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

## WandB
wandb_project: Francois
wandb_entity:
wandb_watch:
wandb_name: Big_boy
wandb_log_model:

## evals
#evals_per_epoch: 4
#eval_table_size: 
#eval_max_new_tokens: 128

## hparams
gradient_accumulation_steps: 3
micro_batch_size: 3
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 2e-5
warmup_steps: 50
weight_decay: 0.0025
## max grad norm
max_grad_norm: 1.0

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
s2_attention:
saves_per_epoch: 2
debug:
deepspeed: ./deepspeed_configs/zero3_bf16.json

#fsdp:
#  - full_shard
#  - auto_wrap
#fsdp_config:
#  fsdp_limit_all_gathers: true
#  fsdp_sync_module_states: true
#  fsdp_offload_params: true
#  fsdp_activation_checkpointing: true
#  fsdp_use_orig_params: false
#  fsdp_cpu_ram_efficient_loading: true
#  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
#  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
#  fsdp_state_dict_type: FULL_STATE_DICT
#  fsdp_sharding_strategy: FULL_SHARD
special_tokens:
  pad_token: <|finetune_right_pad_id|>
  eos_token: <|eot_id|>