control14b.yaml

base_model: allura-org/TQ2.5-14B-Sugarquill-v1
strict: false

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true

# Output and HuggingFace
hub_model_id: NewEden/control-14b-lora
hf_use_auth_token: true
hub_strategy: "all_checkpoints"

wandb_project: huggingface
wandb_entity:
wandb_name: Control-14B


chat_template: chatml
group_by_length: false
datasets:
  - path: Nitral-AI/Creative_Writing-ShareGPT
    type: chat_template
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: Nitral-AI/ARES-ShareGPT
    type: chat_template
    chat_template: chatml
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: NewEden/Claude-Instruct-5K
    type: chat_template
    chat_template: chatml
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: NewEden/OpenCAI-ShareGPT
    type: chat_template
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: NewEden/PIPPA-Mega-Filtered
    type: chat_template
    chat_template: chatml
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: NewEden/Roleplay-Logs-Sharegpt-Ngram-cleaned
    type: chat_template
    chat_template: chatml
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: Nitral-AI/Creative_Writing-ShareGPT
    type: chat_template
    chat_template: chatml
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn

#val_set_size: 0.01
#evals_per_epoch: 1
# eval_table_size:
# eval_max_new_tokens: 128
num_epochs: 2

sequence_len: 8192
save_safetensors: true
saves_per_epoch: 2
logging_steps: 1
special_tokens:

# Quantization
bf16: auto
fp16:
tf32: false
## For LoRA
load_in_8bit: false
load_in_4bit: True

# LoRA
peft_use_rslora: true
adapter: qlora 
lora_model_dir:
lora_r: 128  
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:

## if oom
# lora_r: 64  
# lora_alpha: 32
# lora_dropout: 0.1


weight_decay: 0.02
max_grad_norm: 1.0

warmup_ratio: 0.05
learning_rate: 0.00002
lr_scheduler: cosine
#lr_scheduler_kwargs:
optimizer: paged_adamw_8bit # usually adamw_torch or paged_adamw_8bit

## Batch Size
gradient_accumulation_steps: 8      
micro_batch_size: 1                
eval_batch_size: 1

# Optimizations
pad_to_sequence_len: true
sample_packing: true
eval_sample_packing: false
flash_attention: true
xformers_attention:
gradient_checkpointing: "unsloth"
gradient_checkpointing_kwargs:
   use_reentrant: true
local_rank:
early_stopping_patience:
debug:
special_tokens:
  pad_token: <|endoftext|>
  eos_token: <|im_end|>